Translating Into Success:

Cutting-edge strategies
for going multilingual
in a global age
Volume XI 2000
Robert C. Sprung
EDITOR
Simone Jaroniec
CO-EDITOR
JOHN BENJAMINS
PUBLISHING COMPANY

TRANSLATING INTO SUCCESS

The American Translators Association Scholarly Monograph Series is published

periodically by John Benjamins Publishing Company. Since contributions are
solicited by the Editors, prospective contributors are urged to query the Managing
Editor or Theme Editor before submission. The theme and editor for volume XI
is Language Management, Robert Sprung.
Back volumes of the ATA Series may be ordered from John Benjamins Publishing Company Amsterdam (P.O. Box 75577, 1070 AN Amsterdam, The Netherlands) or Philadelphia (P.O. Box 27519, Philadelphia PA 19118-0519, USA).
Volume I (Translation Excellence, edited by Marilyn Gaddis Rose), Volume III
(Translation and Interpreter Training and Foreign Language Pedagogy, edited
by Peter W. Krawutschke) and Volume IV (Interpreting-Yesterday, Today and
Tomorrow, guest editors: David and Margareta Bowen) are out of print. The
following volumes are available:
Volume II Technology as Translation Strategy. Guest editor: Muriel
Vasconcelles, Washington, D.C.
Volume V Translation: Theory and Practice. Tension and Interdependence.
Guest editor: Mildred L. Larson, Summer Institute of Linguistics (Dallas,
Texas).
Volume VI Scientific and Technical Translation. Guest editors: Sue Ellen and
Leland D. Wright, Jr., Kent State University.
Volume VII Professional Issues for Translators and Interpreters. Guest editor:
Deanna L. Hammond, Washington D.C.
Volume VIII Translation and the Law. Guest editor: Marshall Morris, Puerto
Rico, Rio Piedras.
Volume IX The Changing Scene in World Languages. Issues and challenges.
Guest editor: Marian B. Labrum, Brigham Young University, Utah.
Volume X Translation and Medicine. Guest editor: Henry Fischbach.
Managing Editor: Franoise Massardier-Kenney, Kent State University (Kent,
Ohio). Editorial Advisory Board: Marilyn Gaddis Rose (Binghamton University
NY); Deanna L. Hammond (); Peter W. Krawutschke, Western Michigan
University (Kalamazoo); Marian Labrum, Brigham Young University (Provo,
Utah); Marshall Morris, University of Puerto Rico (Rio Piedras, P.R.) and Sue
Ellen Wright, Institute for Applied Linguistics, Kent State University (Kent,
Ohio).

Translating Into Success

Cutting-edge strategies
for going multilingual in a global age
AMERICAN TRANSLATORS ASSOCIATION
SCHOLARLY MONOGRAPH SERIES
Volume XI 2000

Robert C. Sprung
EDITOR

Simone Jaroniec
CO-EDITOR

JOHN BENJAMINS PUBLISHING COMPANY

AMSTERDAM/PHILADELPHIA

TM

The paper used in this publication meets the minimum requirements of American National Standard for Information Sciences Permanence of Paper for
Printed Library Materials, ANSI Z39.48-1984.

Library of Congress Cataloging Serial Number 87-658269

2000 John Benjamins Publishing Company, Amsterdam/Philadelphia

ISSN 0890-4111

ISBN 90 272 3186 9 (Eur.) / 1 55619 630 X (USA) (Hb.; Alk. paper)
ISBN 90 272 3187 7 (Eur.) / 1 55619 631 8 (USA) (Pb.; Alk. paper)
All Rights reserved. No part of this publication may be reproduced or transmitted in any
form or by any means without prior written permission from the Publisher.
Printed in The Netherlands

Contents
F O R E W O R D . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vi i
I N T R O D U C T I O N . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ix
SECTION I: CROSS-CULTURAL ADAPTATION
1

Microsoft Encarta Goes Multilingual ....................... 1

Bernhard Kohlmeier
2

A d a p t i n g T i m e M a g a z i n e f o r L a t i n A m e r i c a . . . . . . . . . . . . 13

Robert C. Sprung
Alberto Vourvoulias-Bush
3

G l o b a l i z i n g a n e - C o m m e r c e W e b S i t e . . . . . . . . . . . . . . . . . . . . 29

Susan Cheng
SECTION II: LANGUAGE MANAGEMENT
4

What Price Globalization?

M a n a g i n g C o s t s a t M i c r o s o f t . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43

David Brooks
5

Multilingual Information Management

a t S c h n e i d e r A u t o m a t i o n . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59

Cornelia Hofmann
Thorsten Mehnert
6

The Role of Translation in

a n I n t e r n a t i o n a l O r g a n i z a t i o n . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81

Andrew Joscelyne
SECTION III: LOCALIZING THE PRODUCT
7

Localization at Hewlett-Packards
L a s e r J e t S o l u t i o n s G r o u p . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97

Karen Combe
8

Shortening the Translation Cycle

a t E a s t m a n K o d a k . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111

Suzanne Topping
9

Making a Global Product

a t M a p I n f o C o r p o r a t i o n . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127

Ricky P. Thibodeau

vi
SECTION IV: LANGUAGE TOOLS AND TECHNIQUES
10

A Quality-Assurance Model
f o r L a n g u a g e P r o j e c t s . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147

Siu Ling Koo

Harold Kinds
11

T e r m i n o l o g y M a n a g e m e n t a t E r i c s s o n . . . . . . . . . . . . . . . . . . 159

Gary Jaekel
12

Mission-Critical:
T r a n s l a t i n g f o r R e g u l a t e d I n d u s t r i e s . . . . . . . . . . . . . . . . . . . 173

Robert C. Sprung
SECTION V: LANGUAGE AUTOMATION
13

Machine Translation and

C o n t r o l l e d A u t h o r i n g a t C a t e r p i l l a r . . . . . . . . . . . . . . . . . . . . 187

Rose Lockwood
14

Combining Machine Translation

w i t h T r a n s l a t i o n M e m o r y a t B a a n . . . . . . . . . . . . . . . . . . . . . . . 203

Carmen Andrs Lange

Winfield Scott Bennett
15

Language Automation
a t t h e E u r o p e a n C o m m i s s i o n . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219

Colin Brace
A B O U T T H E C O N T R I B U T O R S . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225
A B O U T T H E S P O N S O R S . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231

American Translators Association (ATA)

Officers and Board of Directors
Recipients of the Alexander Gode Medal
Past Presidents of the ATA
ATA Corporate Members
ATA Institutional Members
F O R F U R T H E R R E F E R E N C E . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237

vii

Foreword
The transformations in the language industry following the groundswell of
globalized trade are nothing short of revolutionary. In under 10 years, the
translation and software-localization businesses have evolved from a cottage
industry into the global business imperative. According to the European
Commission, the translation-services market is valued at over US$30 billion
annually and is growing at 1518 percent per year.*
By the early 90s, the major US software companies were averaging seven
languages per each product shipped internationally. They were learning that
the longer it took to enter foreign markets with translated product versions,
the greater the revenue loss. Today, with an emphasis on simultaneous shipment, the larger companies publish products in over 30 languages. By 2005,
Microsoft alone predicts increasing this number to 80 languages.
Ever more companies are beginning to understand the complexity and
commitment behind going globalwhich permeates every facet of an
organizations structure. This includes language-investment strategies, language-processing technologies, translation systems, quality-assurance guidelines, and market education. The Web is only accelerating this trend.
Despite such growth, the language industry is experiencing a severe shortage of talent and information to meet this demand. Major professions like
accountancy, medicine, and law enjoy well-established educational paths,
certification and professional standards, trade organizations, and savvy consumers. Meanwhile, the equally quality-critical language industry is struggling to develop all these trappings of a profession, and to establish standards
and best practices for conducting its business.
Translating Into Success is a landmark publication. It documents the strategic
importance of translation and localization in the global marketplace.
Through case studies and process reviews, readers can learn from the
experts, and benchmark their own operations to those of companies with
leading language solutions. This is a collection of hard-won, real-world
solutions for the language professionalsranging from practitioners like linguists, designers, and software engineers, to project managers and purchasers of language serviceswho are helping fuel this global revolution.
Michael Anobile, founding member and managing director,
Localisation Industry Standards Association (LISA), Geneva
* 1996 MLIS Program, Translation Market Study.

ix

Introduction
These are heady times for the language industry. Riding the wave of globalization, companies are finding that to sell beyond their borders, they must
communicate in their customers language.
A surprising amount of ink is spent on the premise that translation is on the
way out. English is the worlds lingua franca, the argument goes, particularly
after the arrival of the English-dominated Internet. Most people can speak
English, and translation adds large expense and time to product development and international commerce.
The facts say otherwise. Consider these touchstones of a booming language
industry:

The worldwide market for translation and software or Web localization

services is large and growing. Market watchers at Allied Business
Intelligence peg it at US$11 billion for 1999, growing to US$20 billion
in 2004.1

Non-English-speaking Internet users will exceed English-speaking

users by January 2001. According to Forrester Research, business users
on the Web are three times more likely to purchase when addressed in
their native language.2

In fiscal 1998, over 60 percent of Microsofts revenues came from

markets outside the United States. In that same period, revenue from
translated products exceeded US$5 billion. Microsoft executes over
1,000 localization projects a year.

Newsweek, Glamour, Discover, People, and Rolling Stone magazines are

now available in Spanish or Portuguese. Readers Digest reached 600,000
Brazilian readers in 1998, with 1.7 million reading the publication in
Spanish and Portuguese; Glamour en espaol now sells 500,000 copies.
Time, The Wall Street Journal, and CNN have Spanish editions as well.

The worlds population of translators is flourishing. According to a

study for the European Commission, there are practically 100,000
people translating professionally in Western Europe.3 A separate study
estimates 317,000 full- and part-time translators worldwide.4 Translator

1. Allied Business Intelligence, Language Translation, 1998.

2. Rose Lockwood, You Snooze, You Lose, Language International, August 1999.
3. Rose Lockwood, Bigger Than a Bread Box, Language International, June 1999.
4. Allied Business Intelligence.

x Introduction

organizations are thriving: the American Translators Association

(ATA) has 7,000 members in 58 countriesdouble the total of seven
years ago.
Much of the translation performed today is required by regulatory authorities or prompted by liability and safety concerns. An Italian sedan in Madrid
requires a Spanish service manual. An American mutual fund in Luxembourg may require a German prospectus. A French medical device in Tokyo
requires Japanese instructions and labeling.
Much of the remainder is required by a simple marketing analysis. Once a
product exists, translating ittypically a small percentage of the total product costmay open up new and lucrative markets. Twenty years ago, IBM
might have gotten away without translating technical manuals, claiming that
the engineers reading them spoke enough English. Today, most exporters
face local competitorsconsumers in Taipei or Moscow will gravitate
toward the product in their own language, not the one in the strange packaging. Companies are finding that the cost of not translating poses too great a
risk to international sales.
These are some of the foreign-language-related services for which demand
has skyrocketed in recent years:

translationthe core skill of converting text from one language to

another, whether on hard copy or electronically.

internationalizationdesigning a product (e.g., software) so that it

supports usages around the world (e.g., number, date, and currency
formats) and can be easily adapted and translated for individual
local markets.

localizationtaking a product (ideally, one that has been

internationalized well) and tailoring it to an individual local market
(e.g., Germany, Japan). Localization often refers to translating and
adapting software products to local markets.

cross-cultural consultingbooming areas are marketing to ethnic

minorities in a given country (e.g., Hispanic or Chinese communities
within the US) and name verification to ensure that a proposed
product name is effective and inoffensive in other languages. A name is
de facto accessible to all other language communities when it appears on
the Web.

multilingual designtypesetting and desktop publishing in foreign

languages.

Introduction xi

multilingual technology and toolsthese range from automatic

hyphenators in different languages to terminology-management tools
that ensure consistency in translated documents.

machine translationusing a computer to perform a first-pass translation.

A note on usage: companies that specialize in software localization are quick

to distinguish themselves from translation companies. However, those
outside the language business proper are not as likely to draw such distinctions. In this introduction, translation business and language industry
encompass the entire range of services listed above.
W H Y

T H I S

B O O K ?

The emerging language industry is sorely in need of best practicesstandards of excellence which can assist in communications between clients and
vendors, and also aid in the training of those entering the industry. This
book aims to help fill that gap. Each case study strives to capture actual
achievements, not merely theories or opinions, that have broader relevance
in the language industry. Case studies are written by recognized authorities
and leading practitioners, including those who helped pilot the multilingual
effort at Microsoft, Hewlett-Packard, Time Warner, Ericsson, and the
European Commission.
These case studies help answer the fundamental questions in global communications today:

How do we control language costs while maintaining quality?

How do we reduce translation cycle time to improve time-to-market?

How do we develop a scalable language process that can keep up with

our companys growth?

How do we stay on top of language-technology developments?

How do we look at language strategically, to leverage our brands and

products worldwide?

The language industry is largely an industry in name onlya convenient

rubric for a patchwork of sole practitioners, in-house language departments, and small translation companies. Translation is often literally mom
and pop, with husband-and-wife teams of translators operating from
home. Although a recent wave of mergers and acquisitions has created a
half-dozen US$50-million companies, the market remains highly fragmented, with no player holding much more than one percent.

xii Introduction

One result of this fragmentation is a lack of accepted quality standards.

Despite the inherently cross-cultural nature of their work, translators are
not held to any international standards for quality. Since the translator or
localizer in Amsterdam and the one in Tokyo are doing exactly the same
thing, the industry seems tailor-made for such professional standards, particularly given the inability of the general public to evaluate language work.
Some national standards or certifications do exist, but they vary radically
from country to country in both rigor and enforcement. A translator in
Germany or Argentina may meet nationally accepted requirements in education and professional practicethe term publicly sworn translator
means something very specific both to language practitioners and to the
broader public. In many other countries, however, meaningful standards are
either absent or so new and inadequately promoted that consumers are
unaware of them. The customer is thus usually left to fend for himself in
selecting a language solution and evaluating a foreign text.
The educational system certainly plays a role. In Europe, which has been
translating successfully for centuries, educational institutions are more
firmly entrenched than in the US, where translation is largely fueled by a
relatively recent boom in global trade. In the United States, one of the largest purchasers of translation services, a handful of educational institutions
offer degree programs in translation. It is difficult to hold translators to rigorous standards when so many are needed and so few are trained.
Part of the problem, too, is that translation professionals have long had an
image problem. The portrait of translators derived from most reference
books is not flatteringyou might find that the Italians coined the catchphrase traduttore, traditore (translator, traitor). Purchasers of language services
are often unaware of the skill needed to recast text in a foreign tonguethe
typical response to a translation request in many US corporations used to
be: Get a secretary to do it. Translation is often thankless; ask a dozen
marketing managers for their experience, and their only memories will be
of translation errors. A professional translation does not enjoy praiseit
merely avoids criticism.
The tide is turning, however, largely because of the strategic nature of
translation. Billions of dollars are spent each year on translation and related
servicesa tiny fraction of the value of the actual products being sold. Poor
translation means lost revenue through lower acceptance abroad, or even a
potential product recall. Inefficient translation means lost revenue through
slower time-to-market. Given the importance of the multilingual versions
we create, the market is quickly raising the bar on translation quality.

Introduction xiii

On the customer side, professional organizations like the Society for Technical Communications (STC) and the Software Publishers Association
(SPA) have raised awareness of multilingual communications for their members. Leading trade publications increasingly feature serious stories on
translation (not merely rehashing apocryphal gaffes such as Chevys decision
to market the Nova in Latin America, where no va means doesnt
run in Spanish). Consider that Adweek now stresses global branding, with
regular features on pitfalls in international advertising. Crossing cultures and
languages has become a mainstream obsession: consider recent cover stories
in Newsweek (on the booming Hispanic and Spanish-language market
within the US) and National Geographic (on global culture).5
Trade and professional organizations have made progress in industry standards. Many countries have professional translator associations (e.g., the
American Translators Association in the US, SFT in France). The ATA,
sponsors of this book, have made great strides in promoting the profession
of translator, and have taken on the daunting task of developing a standard
for translation certification in the United States. The Localisation Industry
Standards Association (LISA), based in Geneva, helps bring together customers and suppliers in the software business. One case study in this work
features the LISA Quality Assurance Model, which several companies have
successfully implemented to set standards for language projects. Translators
in other countries are striving to organize and establish new quality standards, while broader international organizations such as Germanys DIN are
pushing for international translation standards.6
Educational institutions have broadened their offerings in the language professions. In the United States, the Monterey Institute of International
Studies and Kent State University both offer degree programs in translation
and related fields, and are training a generation of translation producers and
consumers with cutting-edge strategies and techniques. And publications
such as Language International, for which the present writer serves as editor,
strive to provide a written record and forum for this emerging profession.
The language industry has started appearing on the radar screens of major
market-research companies like Forrester Research and Gartner Group. A
few translation market-watchers have emergedwhich implies that there is
a growing market worth watching. Rose Lockwood, who is represented by
a case study in this volume, has helped elevate the discourse surrounding
language professionals, by situating their work within the context of an
5. Newsweek, July 1999; National Geographic, August 1999.
6. DIN preliminary standard 2345; Web site: [Link].

xiv Introduction

InfoCyclethe creation of a companys information and its subsequent life

cycle as it appears in varying forms and versions. Lockwoods Globalisation
study was the first to analyze and quantify the industry formally.7
T H E

C A S E

S T U D I E S :

R O A D M A P

Translating into Success may serve as a casebook for those already working in
the industry, or those seeking to enter it. We have organized the book into
five thematic sections, which give a cross-section of a complex industry in
rapid growth and transition.
Se c t ion I : C ro s s- C ul t ura l A dapt a ti o n

Effective translation bridges the gap between cultures, not merely words.
The case studies in this section show that the most effective way to make a
product truly international is to make it look and feel like a native product
in the target countrynot merely to give it a linguistic facelift by translating the words of its documentation or user-interface.
Bernhard Kohlmeier describes how Microsoft effectively localized Encarta
Encyclopedia into several languages, by creating and adapting content so that
it would appear to be a native product. The foreign versions of this online
encyclopedia actually differ from country to country; the achievement was
not only gathering the appropriate resources to make editorial decisions,
but also developing and managing complex work processes so that the
project could be completed in a very short timeframe. The technical and
procedural innovations in managing this projectwhich included some
33,000 translated articles totaling 10,000,000 wordsserve as an invaluable
source of best practices. Kohlmeier offers memorable examples of the way
Microsoft modified its foreign-language versions to increase sales and avoid
potential embarrassment abroad.
Robert Sprung writes about Time magazines first appearance in a foreign language. The key to success was that the publication did not merely duplicate
its English version, but tailored contents specifically to Latin America. The
case study illustrates the challenges of recreating the voice and style of a text
for a different culturethe same principles apply, whether the text is journalism, marketing copy, or technical documentation. The project required a
robust, scalable work process that could convert 10,000 words every week
into Spanish and Portuguese in little over 48 hours, which required not
7. Rose Lockwood, Globalisation: Creating New Markets with Translation Technology, London:
Ovum, 1995. The current version (1998) is published by Equipe Consortium
([Link]).

Introduction xv

only a top-notch team of language experts, but also cutting-edge communications and publishing technology. The case also illustrates the potential
for international brand leverageTimes series on Leaders for the New Millennium included multilingual content in its English magazine, the Latin
American edition, CNN en espaol, and the Web.
Susan Cheng shows cultural adaptation at work in a small, cutting-edge
e-commerce company, [Link]. When this firm decided to market
golf equipment to Japan via the Internet, they used specialized software and
work processes to localize their e-commerce site. The case study illustrates
the importance of understanding current marketing idioms in a given culture, and applying these in the localized version of a Web site, regardless of
the language of the user-interface. Cheng illustrates how Web translation
differs from translating for traditional media, and gives us an inside view of
workflow and resources used, with copious graphical examples.
Section II: La ngu age Ma nage ment

Smart companies understand that translation is a strategic imperative. It

requires complex planning and analysis, with significant investment of time
and resources to achieve the ultimate goals of lower translation costs, higher
quality, and shorter turnaround time. These cases show how some companies are lifting decision-making for translation and language issues to the
highest levels, rather than treating it as an afterthought or low-end administrative work. Given that the success of their core products is riding on
translation quality, the move seems long overdue.
David Brooks describes Microsofts efforts to bring translation and localization costs under control. Microsoft spends hundreds of millions of dollars
each year on localization, with translation in the critical path to product
delivery. Finding an optimal way to cut costs and time was thus hardly
optional, but a key to survival and success in the global marketplace. In spite
of its mammoth size, the company takes a very entrepreneurial and creative
approach to problem-solving, and the innovations it employed, which are
outlined here, apply to small companies as well as large oneswith implications far beyond the software industry. Brooks, who helped craft the
localization effort at the worlds largest software company, focuses on the
need for true global design of products, and for cost-basis accounting to
rein in skyrocketing localization expense. He discusses the application of
quality and cost metrics to translation at Microsoft, applying Andrew
Grovess dictum, If you cant measure it, you cant manage it. He also
steers the readers gaze toward cutting-edge solutions that have not quite
arrived: Linguistic technologies such as speech recognition, natural-lan-

xvi Introduction

guage-based search engines, and machine translation will redefine the scope
of localization. These technologies are being commercialized, and will soon
become mainstream utilities.
Cornelia Hofmann and Thorsten Mehnert describe language innovations at
Schneider Automation. Multilingual efforts there break down the traditional categories of translation and fit into the broader context of
multilingual information management. The authors argue that the
information of tomorrow is quite different from the documentation of
today. Consider these examples:

Information is increasingly integrated into products (e.g., online help or

on-screen messages).

Products are increasingly information-intensive (e.g., car navigation

systems containing large volumes of translatable data).

The Web is blurring the distinctions between traditional

documentation categories (technical documentation, marketing
literature, or customer support).

The authors advocate abandoning a silo perspective that processes individual documents, and propose an integrated approach that thinks in
terms of recyclable information objects. The real-world examples in this
case study have far-reaching implications for translation in a wired world.
Andrew Joscelyne introduces another facet of the language businesstranslation for public and quasi-public institutions. Organizations like the
European Commission, the IMF, and other international bodies are among
the largest employers of translators and language services in the world.
Many of their problems are specific to the nature of such organizations,
their needs and decision-making structures. Others are similar to the challenges faced by firms, large and small, in the private sector. Joscelyne shows
that international organizations like the Organization for Economic Cooperation and Development (OECD) can innovate as well as their privatesector counterparts.
Section III: Localizing th e Product

This section offers three facets of the complex processes of localization

and internationalization. Internationalization means designing a product
so that it can be easily adapted to local markets. Examples include developing a computer operating system so that it can support complex foreign
character sets like Chinese or Japanese, international date formats, and sorting sequences. A product that takes into account the various possibilities of

Introduction xvii

adaptation around the world is said to be internationalized. Localization is

commonly defined as the process of taking a producthopefully one that
has been well internationalizedand adapting it to a specific locale or target
market or language group (translation is thus a subset of localization). An
example may help illustrate the point: designing an automobile chassis so
that the steering wheel could be installed on either the right or left would be
a case of internationalization. The decision to actually make a given batch
of cars left-steering would be a case of localization.
Localization is most commonly associated with the software industry, since
it is one of the largest consumers of such services. However, we define it
more broadly, to include the adaptation of any good or service to a target
market, since the principles applied in both cases are similar. These three
case studies offer invaluable insights with application far beyond the computer desktop.
Karen Combe outlines Hewlett-Packards approach to localization. HP experienced dramatic growth in international sales, and with it a rising demand
for language and localization services. The case study is particularly interesting in its illustration of the principles of partnership and communication
between a customer and a language vendor. The innovations here point the
way for a truly collaborative and strategic effort in tackling complex language problems. We also see how managing the human aspects of language
projects under intense pressure is as critical as managing the technical complexities. Combe describes how Hewlett-Packard strove to apply founder
David Packers classic management by objective philosophy to localization and translationHPs Information Engineering department
benchmarks its language program based on cost, on-time delivery, and quality. The results are impressive.
Suzanne Topping turns our attention to cutting time-to-market at Kodak,
where she helped implement new approaches in use today. Indeed there
will be time, T.S. Eliot once said. Clearly, he was not writing about the
localization industry, where yesterday often seems too generous a deadline. As soon as word spread that a new Kodak camera was available or soon
to be released in the US, consumers in other markets were impatient for
deliverywhich would be impossible without translated instructions and
associated software. In the not-too-distant past, consumers abroad felt like a
younger sibling receiving a hand-me-down: their brethren in the US would
enjoy version 4, while they had to sit by patiently with version 3. The Web
has changed that for good: since the world knows the instant a new product
or version is available, the goal is to come as close as possible to simultaneous
release of foreign-language versions of a product. Topping shows how Kodak

xviii Introduction

worked toward this goal, which entailed reengineering product development to accelerate translation. Her study is a blueprint for many technology
companies seeking to cut time-to-market.
Ricky Thibodeau shows how localization is a strategic imperative for small
companies no less than for the big players. He reveals the techniques in use
at MapInfo, a smaller but no less potent software developer. His is a lesson
in the sheer economic power of localization: although MapInfo has only
450 employees worldwide (three quarters of them in North America), its
products are available in 58 countries and 21 languages. He provides comprehensive analysis of MapInfos decision-making as it refined its
localization strategy. He shows how MapInfo transitioned from a model
where translation was performed by its distributors to a hybrid model
involving other third parties and a shored-up internal team. Thibodeau provides detailed descriptions of the companys tools and techniquesuseful in
an era when many firms keep such competitive advantages hidden from
public view. Perhaps this is with good reason, given the impressive results
MapInfo has achieved: they cut localization costs by 25 percent, in addition
to reducing cycle time and improving quality.
Section IV: La ngu age Tools an d Te chni qu es

This section covers language technology and process. Advances in these

areas have been impressive, ranging from specialized tools for language professionals (electronic dictionaries and terminology management, for
example) to specialized models for project management, budgeting, and
communication.
Siu Ling Koo focuses on innovative approaches to controlling quality. LISA
has developed a quality-assurance model for language projects, including
standardized methodology for classifying errors. Koos firm has supplemented the model with sampling techniques, so that portions of a language
project can be tested, saving time on QA. If a sample has fewer than a predetermined number of errors, it passes; if not, it fails. She has been applying
the model for a significant period, and illustrates the potential application of
quality metrics to translationa sign of a maturing industry embracing
modern management techniques.
Gary Jaekel describes terminology management at Swedish telecom giant
Ericsson. His case study offers an object lesson in the benefits of controlling
the words we use. Terminology management typically involves a system for
cataloguing, updating, retrieving, and managing terms. Without it, there
would be far greater danger of using incorrect or inconsistent terminology.
Terminology management is critical for reducing cost and turnaround time

Introduction xix

on translation projects: if translators are given standard, approved glossaries

of terms on which to base their translations, quality will be far higher with
greater consistency. Jaekel here outlines how he helped create a terminology
department that ultimately produced a product with over 15,000 core
terms. Terminology is blossoming as its own discipline. Kent States SueEllen Wright, for example, oversees a growing number of graduate students
pursuing this field. Terminology management is hardly an ivory-tower theory: Microsoft recently took a significant share of Trados, a prominent
maker of commercial software for exactly that purpose.
Robert Sprung writes on technology and project-management innovations in
regulated industriesthose with a high degree of oversight or which must
translate text that is subject to strict standards or regulatory audits. Consider
the large volume of text translated for medical, pharmaceutical, automotive,
legal, and financial firmsquality is obviously critical, but pressures continually mount to translate in ever shorter timeframes while containing costs.
This case study focuses on cutting-edge solutions crafted for two Johnson &
Johnson companies, ranging from ISO-compliant language processes to a
custom corporate intranet to manage and control the translation cycle.
Se c t ion V : La ngu age A u tom at i o n

Cant you get a machine to translate it? This question seems to be on the
lips of many vice presidents in charge of international sales and marketing.
Do the math: a company with 10 user manuals totaling 1,000,000 words in
English contemplates a move into 10 foreign languages; the price-tag for
translation could easily top $3,000,000. With the stakes this high, and with
the price of computer processing speed and storage dropping dramatically,
pressures have been considerable to automate the language process.
Using computers to perform translation has been one of the holy grails of
artificial intelligence (AI) since its inception. A machine that can translate is
a machine that can think and write like a person. Easier said than done, it
turned out. Although commercial machine-translation systems have been
around for about 30 years, the vast majority of translation performed in the
world today is still done by those pesky humans. This often surprises American marketing executives who are new to language and translation. Take
the sentence Mary had a little lamb. How do we know that had means
possessed and not ate (as in, John had a little beef and Mary had a little lamb), taught, took advantage of, or, heaven forbid, gave birth
to? The answer is that we dont knowcognitive scientists dont sufficiently understand how our brain places this particular lamb in a nursery
rhyme and not a butcher shop. Since each language has unique mappings

xx Introduction

of words to meanings, the computer can easily end up linguistically mistreating Mary and her lamb. And if the computer has a problem with Mary,
what sorts of problems will it have with legal contracts or technical specifications for a nuclear power plant? Douglas Hofstadter, a pioneer in artificial
intelligence, thinks that the computer might never be able to produce a reliable translation that reads coherently.8
Meanwhile, commercial pressures continue to impel innovations. Some
think part of the solution lies on the output side, in the form of lowered user
expectations. The search engine Altavista pioneered a Web-based machinetranslation (MT) solution by partnering with Systran. Click on a Web
address listed in the search engine, select a language, and voil: a translated
Web page free of charge. Dont expect a polished or even a particularly
accurate translation, however. The system is good for gisting; by having
most key words translated accurately with reasonably understandable syntax,
you can often understand enough of the text to determine its basic meaning
or at least context. If you need a polished translation or one on which the
accuracy is verified, humans will have to get involved.
Another angle of attack is to refine the input. If you can simplify the text to
be translated, the computers problem will be greatly simplified. Rose Lockwood outlines such a solution implemented at Caterpillar over the course of
years. This leading international manufacturer of heavy equipment looked
at the cost and time involved in translation, and helped develop the concept
of controlled authoring. By dramatically limiting the vocabulary and sentence complexity, the company has achieved impressive results with
machine translation.
Carmen Andrs Lange and Winfield Scott Bennett offer another machine-translation case study, this time from Baan Development in the Netherlands.
Baan has successfully combined machine translation with translation memory, software used to store and manage terminology usage on an enterprisewide basis. The authors also demonstrate that much of the challenge of
implementing MT rests as much on the human as the technical side: many
technical writers view their craft as more art than science, and are resistant
to writing in a manner that they consider stilted or to using automated
means in general. For machine translation to be viable in the long term,
those who review the machines output (post-editors) are typically skeptical, and need extensive coaching and training if they are to support
the effort.
8. See Douglas Hofstadter, Le Ton beau de Marot: In Praise of the Music of Language, Basic Books,
1997.

Introduction xxi

Finally, language-technology watcher Colin Brace offers a glimpse into the

European Communitys implementation of Systran, a large-scale machinetranslation system. Brace explains how the EC has gathered terminology
and trained its system over the years, and has effectively customized
Systran for specific types of EC documents. The case illustrates the extensive preparation involved in any MT implementation. There is currently no
all-purpose MT solution on the market; one must invest in technology
and human capital for a specific use. It may take several years and substantial
investment to recoup the investment in an MT systemcautionary words
for that overly optimistic US marketing manager we mentioned above. But
as these three case studies illustrate, it may be worth the trip. And as the
Internet further heightens demand for rapid translation, the coming years
will no doubt bring us other dramatic developments.
A C K N OW L E D G M E N T S

This volume is part of the American Translators Association Scholarly

Monograph Series. The ATA provided invaluable support, particularly
through Professor Franoise Massardier-Kenney of Kent State University.
Professor Kenney was an ideal project supervisor, offering constructive criticism throughout that greatly improved the balance and focus of the book.
Assistant editor Simone Jaroniec, a seasoned language professional, was an
inspirational colleague. She managed the overall production of this volume,
on top of a full workload. She continually held the team to the highest editorial standards, while her keen insight and dedication made the project a
pleasure to work on for the team.
Graphic specialists Ted Assur and Sean Mahoney breathed life into the text,
giving it an attractive design and format. Kristin Padden, who also serves as
the associate editor of Language International, offered her usual inspired editorial assistance. Special thanks to Adela Hruby, Adriana Lavergne, Liza
Loughman, Natalia Mehlman, and Peter Torkelsson for attentive
proofreading.
Thanks also are due Sue-Ellen Wright of Kent State University, Dean
Diane de Terra and Jeff Wood of The Monterey Institute of International
Studies, and Walter Bacak and Muriel Jrme-Okeeffe of the ATA. Each of
these valued advisors plays a critical role in educating the broader public to
translation issues, as well as training the next generation of language
professionals.
Ingrid Seebus and Bertie Kaal of John Benjamins offered critical assistance
in printing and organization.

xxii Introduction

Any errors that remain are my own. We hope you will find valuable insights
in the present volume. A second is already under consideration; if you are
aware of appropriate case studies or wish to correspond on language issues,
do drop a line.
Robert C. Sprung
Cambridge, Massachusetts
October 1999

C R O S S - C U L T U R A L
C A S E S T U DY

Section I: Cross-Cultural Adaptation

Microsoft Encarta Goes Multilingual

Bernhard Kohlmeier
Bowne Global Solutions

T H E

C H A L L E N G E

Since 1993, Microsoft has been distributing the US version of Encarta Encyclopedia, a digital reference title based on a US print encyclopedia. In 1995,
Microsoft was the first to take an electronic encyclopedia into several foreign tongues, including German and Spanish. Bowne Global Solutions was
contracted to carry out the project and provide ongoing updates.
From the start, it was obvious that a plain translation of the US version
would have no chance in the international market, because all content (articles, style, media, and features) was targeted to US users.
From a German or Spanish perspective, local content (locale-specific biographies, historic events, and geographic entities) was missing, while content
developed for the US audience might have confused local customers.
Examples, institutions, laws, and other frames of reference were US-centric.
Further, the manner of presenting facts and judging these facts was problematic for a non-US audience, particularly in articles concerning historical, legal, or political matters.
The challenges of executing the project rested in the following areas:

Content adaptation and creation

Very large project size

Project management

Budget tracking

Media selection and production

Management of updates throughout the project

Paper management

Below is a chart giving a sense of the mammoth project scope. Figures are
approximate.

2 Microsoft Encarta Goes Multilingual

Version 1

Version 2

27,000

33,000

7,200,000/9,000,000

8,200,000/10,000,000

Media Elements

10,000

11,000

Photos and Illustrations

7,000

7,600

Audio Elements

2,000

2,000

Maps/Charts

1,000

1,250

Articles
Words (depending on language)

Videos and Animations

100

110

Web Links

1,500

Bibliographical Entries

3,500

25 articles monthly

25 articles monthly

Yearbook Builder

In this case study, content entities refers to articles and media elements
(videos and animation), as well as maps and charts.
T H E

S O L U T I O N

Project Ma nage ment

A carefully defined project-management procedure, with close monitoring

of all aspects of the project, was crucial for success. Tracking each content
entity with utmost accuracy at any stage was one of the greatest challenges.
Workflow-management tools based on SQL Server and written in Visual
Basic allowed the management team to monitor the project at any time.
Each content and media entity had to be tracked, along with its status, history, class, and workflow. Assignment rules allowed automatic resource
assignment without manual interaction.
With a production time of 18 months (first version) to 10 months (subsequent versions), a project like Encarta Encyclopedia localization seems quite
long. But as soon as the project manager started to break down the total
time available into time available per entity, the schedule became very tight
indeed. Each content entity was tracked based on a detailed and carefully
controlled throughput model. An intranet-based system tracked individual
performance and status of each content entity.
An editor could not reasonably be expected to produce the same volume
each day or week. Content editing and creation is not a conveyor-belt
operation. Besides the human factor (peoples metabolisms differ from day
to day), each article differed significantly in complexity. Some needed
extensive research. Others could be edited without additional investigation.
An article with a few words might require a trip to a library while one of

Bernhard Kohlmeier 3

Deviation in Words

Deviation in Words

multiple pages was straightforward. Project management needed to account

for this; a simulation was created to calculate projected throughput based on
actual, accumulated data. This model allowed for real-time simulation of
the time left in the project, and empowered the project- and content-management teams to adjust to the parameters. This process of schema and correction can be graphically depicted as follows:
Actual

Plan
Review
Target

Time
Throughput to
meet target

Deviation

Corrective action
(add resource)

Time

Making Decisions

The editorial team had content authority for the project, while project
management controlled scheduling, budget, and production. Conflicts
between the major areas had to be resolved throughout the entire project,
requiring creative solutions at all times. What should be done if there was a
last-minute change due to breaking news? How important was the change
to the product and how would it be introduced? Key factors in the decision-making included the trade-off between content requirements and the
risk of jeopardizing a process or build (compilation of an executable software application), or simply the potential for introducing new problems by
addressing another.
Budget Tracking

In a project with several hundred team members, budget tracking of both

internal and external resources is essential to financial success. Tracking of
the project progress and associated costs incurred was supported by custom

4 Microsoft Encarta Goes Multilingual

and off-the-shelf tools. The entire process, from managing costs to generating invoices, can be very time-consuming, and on a management-intensive
project, it can sidetrack the project manager from proactively managing the
project. Therefore the selection of the proper tools to support this process
was critical.
The project team was able to share real-time data on budget spent versus
progress made, including tracking of all expenses. This allowed the project
manager to make strategic decisions based on cumulative experience and
data on the project. Monitoring and managing the quality of several hundred contributors required strict accounting and resource-management
standards. Incoming work needed to be analyzed against quality guidelines,
and contributors often had to rework text to bring it to the required quality
level. A financial process triggering payment had to be linked to final acceptance of a contributors product.
Managing Paper

At some stages of the project, printouts were necessary for both articles and
media. Given the volume of content entities, printouts quickly exceeded
100,000 pages. Not only did the paper have to be tracked and the electronic
version locked, but hard copies also had to be archived in a way that allowed
quick retrieval while preserving a complete paper trail. Revisions marked
on paper had to be proofed against the corrections entered by text-prepping
staff. Another important aspect was the differentiation between a live hard
copy currently being processed and a dead copy used for personal reference
by team members. The need to mark copies as live or dead becomes clear
by imagining the scenario of two paper copies of the same article with different sets of markups. Here it would be difficult to determine the valid
copy, and the older the copy is, the more difficult to determine the revision
history.
We found a solution to this challenge in a place that tracks many different
physical entities: the supermarket. We implemented barcoding on all printouts, and used a print routine that distinguished between system-controlled
and individual printouts. The barcode not only simplified data entry for
resources dealing with paper, it also eased inventory and archiving. In addition, it provided a watermark that could not be easily reproduced. Barcoding significantly reduced the time required to access electronic versions of
paper copies, and also minimized the potential for data-entry errors.

Bernhard Kohlmeier 5
Testing the Results

Testing a project the size of Encarta Encyclopedia brings with it a complex set
of challenges. A single build cannot be completely tested before revisions
are made to the content, as those cycles would produce unacceptable delays
in time-to-market. The solution developed included an ongoing testing and
revision cycle closely integrated with the workflow-management system. At
any given time, multiple revisions were tested against a new build while
editors and media-production staff were working on revisions. Locking
content entities assigned to a specific resource was mandatory to avoid confusion in the workflow and to preclude simultaneous access by different
people or computer processes.
Harnessing Consistency

An encyclopedia project requires the creative contribution of many individuals, which poses the risk of inconsistencies in style and detail. Each team
member has preferences for what is correct, ranging from macro details like
proper representation of historical fact, to micro details such as formatting
and stylistic guidelines.
Here are but a few potential hotbeds of inconsistency:

Spelling of proper names

Format of date ranges

Format of the first paragraph of an article

Guidelines for the type of information to be included

Manner of referencing media in articles

Choice of appropriate media entities for articles

To ensure consistency in these and many other areas, a custom style guide
was created that outlined the content and media requirements in sufficient
detail. The key is sufficient. Although it is easy to create style guides of several hundred pages, the information must be sufficiently concise to be
usable for translators and copy editors.
Constant updates to the style guides and requirements documents reflecting
the latest input by copy editors allowed for corrective action throughout the
project, continuously improving the quality. This loop allowed the project
manager to steer the projects quality in close to real time.

6 Microsoft Encarta Goes Multilingual

Process and
Requirements

Measurement

Corrective
Action

Comparison

Edi t oria l S ubtl eti es

The value of electronic publications rests on their accuracy and timeliness.

With each passing day, many changes in our world potentially influence a
product like Encarta Encyclopedia. New discoveries, changes in geopolitics,
and biographies are a few examples. Historic events must be monitored and
all related content entities kept synchronized with the latest information.
When Yitzhak Rabin was assassinated at a peace rally on 4 November 1995
in Tel Aviv, the news affected not only the statesmans biography but also
many articles about Israel, the Middle East, and international politics. The
workflow-management system efficiently located the related entities and
reassigned them, according to classification, to the responsible editors, and
changed their status within the workflow.
The editorial process needed to include clear style guidelines on how to
edit an article. Editors had to continually monitor the length of the articlepredetermined through the headword list review at the beginning of
the projectwhile balancing time allocated to that task in the overall
project plan.
Editors also had to let go of content entities at a reasonably predictable rate,
otherwise project and editorial management could not effectively plan the
project and monitor throughput. Thus, a key to the projects success was
constant coaching of editorial staff on how to successfully close articles on a
regular basis, and how to perform a self-evaluation of their work using various tools. The support of the editors in monitoring their own performance
and in evaluating their own throughput on an ongoing basis reduced the
project-management overhead.
It turned out to be important to reassure editors that if a headword was
influenced by a news event, they would be able to revisit the headword and
related articles. This is critical for topics subject to frequent updates, and to
the editors peace of mind.

Bernhard Kohlmeier 7
Cu ltura l Ad ap ta tion

Adapting a product to various cultures is very complex and requires deep

and timely knowledge of the target markets. The content of the entire
encyclopedia had to be analyzed based on a thorough review of the headword list, with input from the best subject-matter experts.
Advisory Boards

Two advisory boards were established to review content, write articles, recommend authors and consultants, and suggest new articles and features.
One board comprised 38 academic advisors from Argentina, Brazil, Chile,
Colombia, Mexico, Peru, Spain, and Venezuela, ensuring that this version
of the encyclopedia encompassed the 22 Spanish-speaking countries in the
Americas and Europe. Seven academic advisors staffed the German advisory
board. The advisors had many ideas and suggestions, although it was not
always easy to find a common realistic and practical approach. Direct communication was established either through visits to their respective countries
or regular meetings of the entire board.
Headword List Review

Content balance is a key factor in producing an encyclopedia. Although the

original version provided a great level of detail in some areas, the team had
to ensure the proper balance of headwords for the target locale. Numerous
topics and categories needed expansion. For example, a Spanish or German
user expects a more detailed treatment of European and national history
than is provided in the US version.
Editorial management and subject editors reviewed the headword list. During this phase, decisions were made whether to keep or omit certain articles. The next step was to decide on a process for determining the proper
course of action for an article associated with a given headword: adaptation
of content, completely new material, or a full or partial translation.
For all articles authored locally, length and content requirements were
defined. The article was then written by authors specialized in the respective subject area and edited internally. Translated articles were classified
according to the magnitude of the revisions necessary.
All the factors mentioned above were reflected in throughput projections,
assignment rules, and revision cycles. Appropriate approval processes for
articles of high cultural sensitivity in the respective market were also
defined.

8 Microsoft Encarta Goes Multilingual

Fact-Checking and Research

There are many sources for facts in the world, many of which differ significantly from each other. Geopolitical issues are referred to in different ways
depending on the local standpoint, for example. These phenomena influence large parts of a product such as Encarta Encyclopedia, which necessitates
extensive fact-checking throughout the editorial effort. The goal, from the
outset, was that the product should represent the generally accepted view of
the locale that would be found in the most widely accepted reference material. In numerous instances, editors performed additional research to gather
information that could not be translated in a straightforward manner. A
good example is the way in which titles of movies and books are referenced
in a German encyclopedia, where the title of the original work is given
along with the local title. As the local title is only rarely a literal translation
of the original, research was needed to identify the correct translation: the
Spielberg film Saving Private Ryan is entitled Der Soldat James Ryan in German (the soldier James Ryan), for instance.
Tai lo r in g C o nt e nt t o L o c al e

Areas of high cultural sensitivity required specific attention by the editorial

team. A certain fact or name used in one language and locale may be inappropriate or even insulting in another. Geopolitical issues are particularly
sensitive due to their close relation to a nations or peoples history and sense
of identity.
Additional coverage of content areas sometimes required additional media
work. The introduction of the euro was deemed so important in Europe
that a link in the Timeline had to be added in the German product, to cite
one example:

Added to the
German version
of the
Timeline

Bernhard Kohlmeier 9

A typical case of differing geopolitical perspectives comes from the Spanish

Encarta Encyclopedia. Although the headword Falkland Islands is used in
English-speaking locales and is acceptable in the US, it cannot be used in
the Spanish product due to its unacceptability in Argentina where the product is sold. The Spanish product refers to the island group as Islas
Malvinas.
The same principle applies to disputed geographical areas such as borders,
where it is often necessary to consult numerous subject-matter experts and
even governmental institutions in the target countries. Historical events are
another major source of problems, as different cultures will often have varying viewpoints on methods of classifying reality.
Depending on the target market, adaptation of wording is sufficient in some
cases. In others, completely new articles must be created, resulting in new
content relationships. This approach effectively leads to a new hyperspace of
content relationships specific to the locale. The content that the user reads
is different, and the approach to linking and associating content entities is
redefined.
An example of this challenge is the case of German reunification. Although
the US version of the product covers the topic in several articles, it does not
exist as a separate headword. Given the importance of the topic to Germans, a separate headword was created that added substantial information
about the political situation leading to reunification. This included videos
and a new map of the occupation zones following the Second World War as
a result of the Potsdam Conference.
Screenshot of German-Adapted Content

10 Microsoft Encarta Goes Multilingual

Another example is Bovine Spongiform Encephalopathy (BSE), or

mad-cow disease. The scientists who discovered it, Hans-Gerhard
Creutzfeldt and Alfons Jakob, are German and their biographies are separate
headwords in the German version. Creutzfeldt and Jakob were considered
sufficiently covered under one headword in the BSE article of the US
product.
Here are some additional changes made in the German version:

Due to fundamental differences between US common law and German

civil law, most law articles were replaced and several new headwords
added. Examples include: Role of Jury, Role of Leading Cases, Role of
Cross-Examination.

German sign-language conventions differ from US standards. The

relevant media element had to be completely modified, as illustrated
below:

Gut

Schlecht

Because of the extremely sensitive nature of articles on National

Socialism and Adolf Hitler, related content had to be heavily adapted to
the German market. Special attention was given to the latest research
and current discussion in Europe and Germany. Given the critical
nature of this period of German history, articles had to added due to
their significance to the locale. Examples include: Arbeitsdienst (forced
labor), Arier (Arian), Bcherverbrennung (book burning), Dolchstolegende
(the belief that the German army of World War I had been stabbed in
the back by the nations civilians), Entnazifizierung (denazification),
Ermchtigungsgesetz (Enabling Act), Roland Freisler, Gleichschaltung
(nazification), Hitler-Stalin-Pakt.
For the German user, the depth of the original US articles concerning
National Socialism was insufficient (background, connections,

Bernhard Kohlmeier 11

developments, ideology, etc.). By contrast, articles on American history

(biographies, events like the American Civil War) were too detailed.
Editorial changes needed to reflect the needs of the locale with an
extremely thorough review process to avoid the slightest ambiguity in
articles covering such topics.

In the subject area of Biology, several locale-specific animals had to be

added. Examples include: Black-Throated Jay, Common Otter, Grass
Frog, Greyleg Goose, Common Viper.

C O N C L U S I O N

Globalizing a multimedia product with over 30,000 articles and 10 million

words presented challenges in project management and cultural adaptation
that were new to the parties involved, as well as to the language industry in
general. In a sense, the project itself entailed the adaptation of the language
vendors process and corporate culture, in order to deal with the highly topical content produced under tight deadlines. In the end, the product on
which we collaborated and innovated would itself play a role in educating
people around the world on issues of history, science, and culturean
accomplishment which we could all find satisfying and fulfilling.

Note: Images from Encarta Encyclopedia are Microsoft Corporation.

C R O S S - C U L T U R A L
C A S E S T U DY

21

Adapting Time Magazine

for Latin America
Robert C. Sprung
Harvard Translations, Inc.
Alberto Vourvoulias-Bush
Time Latin America
O V E R V I E W

This case study examines Time magazines first foray in a foreign language
a weekly news supplement in Spanish and Portuguese for Latin America.
We explore the business decisions behind the move, and outline the innovations employed by Time, Inc. and its language partner, Harvard Translations. The challenge was unique for these reasons:

Time hoped to preserve its signature style in the foreign-language copy,

and not merely provide an accurate translation.

Harvard Translations needed to provide a robust, scalable work process,

converting 10,000 words every week into Spanish and Portuguese in
little over 48 hours.

The tight timeframe required novel application of technology, with

remote access to sophisticated publishing software and electronic
download of final pages by Times Latin-American partners.

This collaborationwhich has produced over 75 weekly issues to dateis a

success story that yields intriguing and profitable insights for clients or language-service providers, both within the publishing sector and beyond.
T H E

M A R K E T

O P P O R T U N I T Y

Purveyors of informationwhether print-, broadcast-, or Web-basedare

going global no less than their brothers who hawk hard goods. Some have
broadened international distribution by expanding English-language readership: The Economist and The International Herald Tribune have long been
staples for innocents abroad and for natives wishing a distinctive global perspective. What is new is the intensity with which the major media are going
multilingual. Readers Digest publishes in many languages, National Geo-

14 Adapting Time Magazine for Latin America

graphic has just launched a French edition, and translated Web sites for major
publications are gaining momentum.
With the US newspaper market sluggish, Latin America presents particularly fertile ground for international publications, according to a recent feature in The New York Times. Newsweek, Glamour, Discover, People, and Rolling
Stone are now available in Spanish or Portuguese, cashing in on a thirst for
their distinctive US spin. Says Fortune magazines Gary Neeleman, People
in Ecuador want to read about Bill Gates.1 Each Readers Digest reached
600,000 Brazilian readers in 1998, with 1.7 million reading the publication
in Spanish and Portuguese; Glamour en espaol now sells 500,000 copies.
The Wall Street Journal chose a publication vehicle that would ultimately be
chosen by Time as well: a newspaper supplement in local newspapers. The
Journals Portuguese supplement has appeared since 1997, and its Spanish
version, which reaches 2.2 million, since 1994. The investment in Spanish
and Portuguese is particularly efficient: It is the only region in the world
we can cover entirely in two languages, says Hugh Wiley, president of
Times Latin-America division.2
Time Warner is no stranger to multilingual publishing: it offers People en
espaol, Fortune appears as a Spanish and Portuguese newspaper supplement,
and CNN broadcasts in Spanish (with an associated Web site). Time magazine itself has long had local editions in Asia, Europe, and Latin America.
These English-language publications offer differing editorial content and
balance, while preserving the magazines look-and-feel. Cautious about
potential impact on its flagship magazine, the company decided to walk
before it ran, and adopted the supplement approach, partnering with leading newspapers throughout Latin America. Time would supply content
final pages in the local languagewhile the partner would provide the
printing and distribution. The partners would share responsibility for ad
sales. Times partner network includes:
Ambito Financiero
Folha de So Paulo
Reforma/El Norte
El Tiempo
La Tercera
El Nacional
El Comercio

Argentina
Brazil
Mexico
Colombia
Chile
Venezuela
Ecuador

1. American Publishers Add Readers in Booming Latin America, The New York Times, May
11, 1998, first business page.
2. Ibid.

Robert C. Sprung & Alberto Vourvoulias-Bush 15

El Diario de Hoy
Gestin
Listn Diario
Siglo XXI
La Prensa

El Salvador
Peru
Dominican Republic
Guatemala
Panama

Sample Supplement Covers in Spanish and Portuguese

16 Adapting Time Magazine for Latin America

CHALLENGES AND SOLUTIONS
Ma rke t i ng Ch al le nges

Times local-language supplement is managed by Time Latin America, a

division of Time, Inc. based in New Yorks Time & Life Building. Hugh
Wiley, president of the division, is responsible for overall business and marketing decisions for the supplement. George Russell, editor of the Englishlanguage Latin-American edition, oversees content and production of the
local-language supplement. Alberto Vourvoulias-Bush, a published author
and former college professor originally from Guatemala, directs the supplements day-to-day operations in Spanish and Portuguese.
The personal involvement of these decision-makers illustrates the importance Time gave the local-language supplement from the start. They
decided not only on the overall direction and positioning of the supplement, but also on editorial content and its translation into Spanish and Portuguese. This contrasts markedly with many publications where translation
is perceived as a more mechanical or even secretarial endeavor.
Once the decision was made to bring Time to Latin America in local languages, the first challenge was in designing a publication. Each week, Russell and Vourvoulias-Bush select articles to be translated from the story list
of Times Latin-American edition. These stories are then re-edited, due to
the different size and format of the Spanish and Portuguese publications.
The story selection reflects the magazines commitment to covering Latin
America and the weeks top world and US news stories. Care is also taken
to preserve the balance of topics characteristic of Time and to appeal to a
broad range of readers with varied interests. Included in each issue are articles on business, discoveries in science and medicine, and trends in culture
and the arts. Through feedback from its newspaper partners in Latin America, in fact, Time learned that readers were particularly keen on the Times
perspective on social and cultural trends in addition to relying on it for
more traditional news stories.
In an age of information overload on cable television and the Internet,
Times most valuable asset was its analysis and distinctive attitude. Diluting
or losing these in translation could impair the companys global brand and
core product. In an increasingly competitive global market for information,
having an edge through superior story selection, content, and style, was
considered to be worth the investment.

Robert C. Sprung & Alberto Vourvoulias-Bush 17

Th e B el ly o f th e B ea s t

Time first sought a translation source. Given the need to assemble a team of
a dozen linguists and project managers working within a tight timeframe,
Time opted for a translation company rather than freelancers.
Finding linguists who could respectably convert movie reviews by Richard
Corliss or book reviews by John Skow into idiomatic Spanish or Portuguese
was a tall order. Consider the inherent difficulties in recasting these sample
English passages in a foreign language. One such article would be enough
to test a seasoned translator. Translating 10,000 words, with an unpredictable mix of subject matter and style, would prove a daunting linguistic
hurdle:
Who or what owns a public life? Surely not, in an age of celebrity babble,
the public person who lives it. And not some over-delicate concept of
historical truth.
A distinctly sleazy case in point is The Rich Mans Table, an expertly
written novel by Scott Spencer, that rips off the life of Bob Dylan. Most
readers will recall that Dylan was a scraggly haired harmonica player with
an edgy voice and an edgy mind, a Jewish kid from Minnesota who
changed his name from, lets see, Zimmerman? The character who
dominates Spencers novel was born Stuart Kramer, transformed himself as
Luke Fairchild, and came from the Midwest, not specifically from
Minnesota. Otherwise, the fit is exact.
The translator who insists on finding the right translation for lets see is
doomed from the start. In the memorable guideline of editor Alberto Vourvoulias, Let the translation sparkle where it wants to sparkle. This might
entail recasting a hip, staccato syntax into a periodic style more appropriate
to Spanish, or leaving out altogether English quips or catch-phrases that
would deflate upon translation.
Especially challenging were passages where you could hear America singing. Latin partners expressly wanted the latest perspectives from the States,
and the localized version was not to dilute or dumb down the original
text, which often brimmed with Yankee sound and fury:
Youre driving from the office, the fuel tank is nearing empty, and so is your
refrigerator at home. For Americans ravenous of appetite but starved of time
(and in need of an oil change), a quick turn into the Chevron station off
Interstate 680 in San Ramon, Calif., is the answer. Thats right,
Chevron, purveyor of premium gasoline, is serving fresh panini, threecheese pesto and double espressos along with its usual selection of octanes.

18 Adapting Time Magazine for Latin America

Or consider these verbal pyrotechnics on consumerism in the US:

And now for a moment of alarmism: imagine a high-school geometry class
getting started in Anywhere, USA, sometime in the pre-asteroidal future.
Students, outfitted in mandatory Nike uniforms and sluggish from lunch at
the Pizza Hut-only cafeteria, lumber into a classroom where the teacher
bears a disconcerting resemblance to the Nissan man. He begins his lesson,
using a Dorito, one of the schools now required visual aides, to try and
explain the Pythagorean theorem.
Or this hip review:
But the largest fun lies in the other characters: jut-jawed Kent Mansley, the
funny-dumb government agent who has bought into the whole duck-andcover thing; Dean the beatnik junk sculptor whose cool helps thwart Kents
heat; Hogarths mother, an old-fashioned, benignly clueless sit-com mom.
The linguistic dilemma: the more fluent a native Spanish or Portuguese
translator, typically the less grounded he or she is in the network of idioms
and associations of the source text. More concretely: the translator who can
write the equivalent of benignly clueless sit-com mom in Spanish probably wouldnt quite get it in English.
Or consider these winners and losersquick thumbs-up or thumbsdown to people in the news. Many references require explanation for
non-English speakers, yet the layout allows no text expansion. Translators
would thus need a serviceable telegraphic style for this type of piece. The
first example features two basketball players charged with sexual harassment
(they played for the Washington Wizards; the untranslatable headline is rendered in Spanish as crying foul); the second spotlights a judge in the
Clinton case.
WAYWARD WIZARDS
Chris Webber and Juwan Howard
investigated; this pair needs more focus,
fewer fouls

MARCANDO FALTA
Os jogadores de basquete Chris Webber
e Juwan Howard so acusados de assdio
sexual

NORMA HOLLOWAY
JOHNSON
Exec-privilege judge insists on keeping
hearings closed.

Jueza en el caso Clinton insiste en

audiencias a puerta cerrada.

Robert C. Sprung & Alberto Vourvoulias-Bush 19

JON AMOS
Through sheer willand wheel
Voluntad de hierro: el atleta britnico en
power, disabled British athlete sets world silla de ruedas marc rcord mundial de
endurance record
resistencia
OLIVER STONE
Busted! Stone-d director agrees to enter
rehab to avoid jail time in drugs-anddrink case

Detngase! Director acude a

rehabilitacin para evitar encarcelamiento
por consumo de drogas y alcohol

Headlines offered a special challenge. Many publications have separate

headline writers, and for good reason. The best ones are masters of wordplay and allusioncrossword aficionados with a poetic touch. Times original headlines are often untranslatable; a faithful translation would yield
gibberish or unsightly copy given the pages geometry. The New-York
team of copywriters would thus style fresh headline for the translated piece.
These actual headlines illustrate the problem:
The Man Who Would be King Kong (a piece on Serbian strongman
Milosevic, in a multiple allusion to Kipling and a gorilla).
Not Fonda Georgia, Jane? (a short on Jane Fondas comparison of US
state Georgia to a third-world country; the piece ended by
mentioning her apology, followed by the equally untranslatable
parting shot: Peachy).
Read My Zipped Lips (a blurb on George W. Bushs silence on
whether he took cocaine as a youth, weaving in an allusion to his
fathers signature read my lipsa close translation would be
laughable).
Conve r ti n g t h e Tex t

After extensive testing of some dozen vendors, Time chose Cambridge,

Massachusetts-based Harvard Translations, Inc. as its translation supplier.
HT assembled a team of language experts and project managers. It was hard
enough finding an individual translator to convert a Corliss review into
Portuguese. The broader problem rested in finding a team of linguists who
could translate quickly and consistently over time. Technical translatorsthose
who might excel at medical or computer manualsonly very rarely can
double as literary translators who can recreate an authors voice in a foreign
language. But technical translators are the ones used to delivering high volumes on tight deadlines. Literary translators typically work on books, often
with several months to create a first draft.

20 Adapting Time Magazine for Latin America

Further complicating the issue was the breadth and variety of stories. A
Corliss or Skow review required a translator with an ear for the music of
language who was also well versed in current English usage. But these literary pieces made up a small albeit critical part of the publication. A given
issue would feature longer pieces on politics, economics, or science. Sample
cover stories included: money laundering, e-commerce, Latin music, biodiversity, gun control, the shroud of Turin, and Sinatra. And breaking news
would routinely provide a last-minute shift in subject matter.
Given the nature of the news business and editorial process (Time often goes
down to the wire late on Saturday), HT never knows the exact content
each week until it receives materials a mere 48 hours before they are due.
One week might feature a technical piece on medicine, while the next
might have an equally technical piece on monetary policy.
HT found it optimal to recruit and train a rotating team of some eight linguists, with a range of technical expertise and writing styles. But they all
had to exhibit certain baseline skills, including:

ability to work reliably under intense time pressure

several areas of specialization (one team member, for example, is a

medical doctor with a degree in literature and strong interest in
classical music)

strong research skills, with the ability to quickly scan sources like the
Web for terminology, but with the intelligence to evaluate the
reliability of such sources

strong fluency in English, with a keen ear for idiom

a professional, fluid style in the target language, with a penchant

for brevity.

A brief aside on the topic of text expansion. Most who have dealt with
translation tell you to prepare for translated text to expand as much as 20 or
30 percent when making the journey from English. There are a variety of
reasons for this, including:

The explicative nature of translation. Translators do not translate words,

they translate ideas. With a culture-bound text like Time, much of the
translators task rests in explaining or repackaging the message for a new
audience. There might be no easy Spanish equivalent for mosh pit;
an effective translation might require a half-dozen Spanish words to get
the point across.

Robert C. Sprung & Alberto Vourvoulias-Bush 21

The syntax of Romance languages. English noun strings (defense

department spokesman) require prepositional phrases in most
Romance languages (the spokesman for the department of defense).

Translator sloth or greed. Many translators go for the path of least

resistance: it is easier to write more words than fewer. Technical
translators (unlike many of their literary counterparts) are typically paid
by the word, which further incentivizes verbosity. Due to the fact
that garners 50 cents; because will only get you a dime.

While we originally budgeted for text expansion of 15 percent, we found

that translations did not have to be much longer than their English counterparts. The actual expansion factor would depend on the nature of the text
to be translated. Omitting any technical detail on a medical or political
piece is unacceptable, and greater expansion is inevitable. Translators have
greater liberty with a cultural piecepreserving every Corliss alliteration or
allusion would produce a stilted or incomprehensible version in Spanish.
Movie or book reviews could often occupy the same amount of space as the
English original.
Brevity was critical to preserving Times voice. Translators working under
pressure, without the luxury of putting their work aside and reviewing it
another day, would require a steep learning curve to pre-edit their own
translations and omit needless words.
Harvard Translations put each article through three processes:

Translation by a subject-specialized language professional (the week

when the shroud of Turin appeared on the cover, the lead translator
happened to have a degree in religious studies; it didnt always work
so smoothly).

Editing by a second native speaker. This entailed a close review of the

text against the original.

Mechanical review by a linguist to verify completeness of text and

accuracy of names and numbers. This stage also provides feedback on
the quality of the prior two stages.

HT developed detailed style guides and glossaries of key terms to heighten

consistency. We also assembled a library of standard journalistic reference
works (such as the official style guide of the Folha de So Paulo, Times partner in Brazil). Our custom project style guide, which was available electronically and totaled some 20 pages, included sample translations and common
pitfalls. Here are a few samples:

22 Adapting Time Magazine for Latin America

North America Free Trade Agreement (NAFTA) is translated as

Tratado de Libre Comercio (TLC).

American is norteamericano; use estadounidense for variety.

The symbol for ordinal numbers (1st, 2nd, etc.) in Spanish and
Portuguese must be consistent. On a Mac, type option-9/0 for
underlined superscript a/o. Do not use the degree symbol ().

HT developed a standardized translation process with these features:

Documented procedures, in compliance with the companys ISO quality

systems; these include a paper trail of who did what at each stage of the
project. In a given week, there might be 30 project components (15
pieces in two languages); since each goes through five or more
linguistic or administrative actions, 150 discrete actions must be
monitored and managed.

Regular feedback loops to all project team members, including formal

project recaps with suggestions for improvement.

A proprietary project-management system that tracks and invoices projects.

Some key lessons we learned from the translation:

It proved critical to have an English native speaker present at all times;

when there was a breaking news story, even the most savvy translator
could misinterpret a trendy turn of phrase.

A translators apparent subject specialization was no guarantee of

quality. We found a number of candidates with degrees in the sciences
who would make basic translation errors, or who would provide literal
or even inaccurate translations of key technical terms. Seasoned
technical translators with subject specialization are much more
successful than technical experts who are new to translation.

Excellence on translation tests is no guarantee of future performance.

Those who emerged as star translators did not necessarily appear so at
first. Commitment and a strong work ethic proved as important as
linguistic skillslate-night and Sunday work were the norm, and a
nine-to-five attitude would have spelled disaster.

Poli shing the Text

After receiving translated copy from Harvard Translations, Time has little
over 24 hours to do the following:

proof the copy

Robert C. Sprung & Alberto Vourvoulias-Bush 23

make corrections to the text where last-minute changes modified the

original article after it was sent for translation

fact-check questions raised in translation

polish the copy

edit the articles to fit their redesigned layouts

write headlines and photo captions

process the pages

post them for transmission to Latin America.

The key to making the process work, under these extreme time constraints,
is editorial teamwork. There are two in-house editors per language. Each
must be fluent (linguistically and culturally) in both English and either
Spanish or Portuguese. Working together with the top editor, they share
duties for all aspects of the editorial process, from research and fact-checking to copy-editing and proofreading the final pages.
Logi stica l a nd Te chn ica l Ch al le nge s

The graphics below illustrate the overall project timeline each week, along
with the roles of key players.

24 Adapting Time Magazine for Latin America

Who

What

When

Time

sends text to Boston

Saturday
midnight

HT

delivers translated, edited text to NYC

Monday
morning

Time

style copy, writes headlines, finalizes

graphics and layout

Tuesday
evening

Partners in
can download PDF (Acrobat) files of
Latin America final pages

Tuesday
evening

HT used technology that proved invaluable, notably its proprietary projectmanagement system. Time provided other key technology, without which
rapid turnaround would have been impossible. Key components were a
high-speed data network linking all parties, and Quark Publishing System
software (QPS). QPS includes QuarkXpress, the popular desktop-publishing and layout application, but is customized for a publishing environment.
QPS was largely developed for Time, Inc., and it serves as the workhorse
for Time magazine (other large publications, including Readers Digest, use it
as well). It has all the features of QuarkXpress, with these key additions:

The key working unit of text is the story. Stories can be a long feature,
or even a word or phrase (e.g., an individual photo caption or story
byline). Anything that needs to be tracked separately may be defined as
a story (the person writing the caption or headline may be different
from the person writing the article).

QPS is a networked application, with identified users granted access

levels and passwords. One access level might permit the user to erase a
file or change access levels, for example.

Stories go through a customizable cycle, and could be routed in this

sequence, for example: writer -> editor -> translator -> proofreader.
No two people can work on the same story simultaneously; stories are
checked in and out, as in a library.

The system is real-time: the most recent version checked in by an

author can be viewed immediately by anyone with suitable access.

QPS has a galley view, which allows the author to see plain text (as in a
word-processor), with tools to control word- and line-count. The
system also has an annotation feature; anyone in the process can add
queries or comments. Annotation also lets editors squirrel away extra

Robert C. Sprung & Alberto Vourvoulias-Bush 25

text, in case a story later runs short. A redline feature retains every past
version of a story (i.e., its final state after every action is completed).
This ensures a virtual paper trail (who changed what and when), and
allows comparison between any two versions. (A paper trail has become
critical in journalism, given the potential for lawsuits.)

A page-layout view allows the author to see the story in graphical

context at any time. Since images belong to the layout and not
individual text stories, designers can work on graphics while the author
writes or edits. Further, there is no danger that an author can alter
graphic content.

The high-speed network meant that anyone with access to QPS could
work in the same virtual office. This allowed instant transfer of files
between Time and translators. It also allows projects with intense deadline
pressure to take advantage of time differences around the world, if a global
team were assembled.
Times New York office also used Quark Passport, a variant of QuarkXpress
that includes hyphenation for a number of foreign languages. None of these
electronic hyphenators is foolproof; further, Time has particularly high aesthetic standards when it comes to line breaksno machine can replace the
eye of a copy editor.
E VA L UAT I N G

T H E

R E S U L T S

Time has been producing its weekly Spanish and Portuguese supplement
for over a year and a half.
From a business perspective, the venture is a major success. Readership is
climbing, with a number of major new newspapers in Latin America signing up as partners (this is a major vote of confidence on their part, as they
must invest in the production and also sell a substantial part of the localmarket advertising). Readership currently totals 1,347,000 per week.
To determine the level of acceptance of the supplement, Time hired an
independent third party to conduct a reader survey. Based on 2,000 completed questionnaires, the research firm found:

86 percent of the newspaper subscribers saw or read the supplement.

Of those, 48 percent read it every week or almost every week.

86 percent feel the supplement is a very valuable or somewhat valuable

addition to the newspaper.

40 percent spend nearly 30 minutes reading the supplement.

26 Adapting Time Magazine for Latin America

25 percent spend 3060 minutes reading the supplement.

44 percent read all or nearly all of the supplement.

36 percent read between half and three quarters of the supplement.

The supplement had good pass-through readership, with each copy

reaching an average of three readers.

Time has positioned the foreign-language supplement as part of a larger

program for Latin America. Consider the following synergies on a recent
campaign:

Times Latin-American English-language edition features a series on

Leaders for the New Millennium, which focuses on 30- and 40somethings who will shape Latin America in the coming years.

The supplement offers Spanish and Portuguese versions of the feature,

bringing it to over a million additional readers.

CNN en espaol produces television programs featuring leaders profiled

in the series.

CNN and Time reinforce the Leaders campaign with local-language

Web pages.

Time has been able to sell this package to corporate sponsors such as Compaq, BellAtlantic, and Mastercardone-stop shopping for reaching a broad
segment of the Latin-American market. In the survey referenced above,
93 percent of respondents agreed with the statement, I would like to see
other editorial initiatives similar to Leaders in the future.
The screenshots below come from a full-color and sound CD-Rom demo
that introduces the Leaders feature. The first screen shows the cover of the
Latin-American edition. Upper right is the supplement in Spanish, and
below it is a video clip from CNN en espaol. Note that corporate sponsors
also appear in this demo CD.

Robert C. Sprung & Alberto Vourvoulias-Bush 27

From the perspective of Harvard Translations, Times language vendor, the

program has been a major success as well. Our contract has recently been
renewed, and we see the relationship as a model for a collaborative relationship. In an industry where the typical question asked of translators is, how
much do you charge per word, it is immensely gratifying to work on a
project where quality mattersand the client knows the difference.
It is only fitting to end with a word of thanks to the many team members
that make the project a success week in, week out. Far from shrinking at the
latest twist of fate that will determine this weeks cover storyand their
weekend schedulethese linguistic gymnasts seem to thrive on it. The
team is also proud that, in their own way, they are helping bring news and
insight to an engaged audience around the world.

C R O S S - C U L T U R A L
C A S E S T U DY

Globalizing an
e-Commerce Web Site
Susan Cheng
Idiom Technologies
I N T R O D U C T I O N

Enabling a sophisticated corporate or e-commerce Web site to be multilingual and multiregional is a complex and subtle undertaking. A Web site is
technology-based and interactive, with text continually changing. Standard
approaches to translating text and software are inadequate for this new
information vehicle.
This case study profiles the globalization of a dynamic Web site, from conception to implementation. We showcase an e-commerce site that was
adapted from English into Japanese, crossing numerous linguistic, cultural,
and technical hurdles in the process. We also profile new language technologies and processes tailored to the World Wide Web.
WE B - S I T E

O V E R V I E W

[Link] is a midsize e-commerce site that sells custom-made golf

clubs via the Internet. First published in 1996, the site is marketed as the
Internets largest retailer of custom-built golf equipment. Some basic facts
about [Link] include:

It features standard e-commerce site layout, including a common

navigation bar and multiple product listings.

Content is displayed within a template.

The entire site is based on proprietary, noncommercial design and

infrastructure.

Product names and prices are pulled from a live SQL database.

Content text is updated or added daily (e.g., product names, product

descriptions).

Functional features are updated or added weekly (e.g., a gift-certificate

suggestion center).

30 Globalizing an e-Commerce Web Site

Given golf s popularity in Japan, Chipshot recognized the opportunity to

expand sales there by making [Link] fully functional in Japanese.
P R O J E C T

E V A L U A T I O N

The project of globalizing [Link] for Japanese viewers was evaluated

according to: 1) the front end (what the user sees); 2) the back end (the
technology and processes behind the scenes); and 3) long-term Web-site
development. The table below summarizes the goals identified for
[Link] globalization:
Linguistic

translated text and graphics

converted/formatted prices and dates
consistent look-and-feel

Front End
Cultural/Marketing

brand promoted for target country

well-advertised site

Technical
Back End
Managerial
Long Term

foreign text correctly displayed

character encoding issues addressed
translation integrated with development
sites synchronized on ongoing basis
changing functional features
future Web-site needs

Fron t En d

Chipshot outlined a set of goals for how a globalized [Link] would

look, feel, and function in Japan. The existing site design was reviewed to
identify front-end initiatives that would make the site a viable revenue-generating site in this new market.
Linguistic

A Japanese [Link] needs to contain translations and cultural adaptations of the English sites textual content, including introductory remarks
on the home page and components of the navigation bar. Given the commercial nature of the site, language and presentation were critical. Almost
all text and graphics would be translated in a culturally targeted way that
reinforced the sites marketing message and merchandising functions.
Globalizing the core shopping features required that product prices be displayed in yen with current currency conversions. Prices on the Japanese site
would be converted directly from US-dollar counterparts on the English-

Susan Cheng 31

language site. Instead of presetting the conversion rate, Chipshot decided to

pull it from a live source.
Cultural/Marketing

It was important to balance two potentially competing considerations: the

site should be culturally appropriate for Japan, but at the same time should
have a look-and-feel consistent with the English site. Colors, images, and
page-by-page format should be the same, to the greatest extent possible.
Chipshot marketing executives aimed to preserve a consistent visual and
functional format to promote brand awareness and to foster general visitor
familiarity and comfort during page navigation. However, recognizing that
the browser standard for presenting Asian-language Web-site material may
change in the future (e.g., Asian characters may be displayed vertically
instead of horizontally), the company wanted to retain the flexibility of
changing page-by-page formatting of the Japanese site.
Chipshot decided not to translate certain product names, product descriptions, and marketing slogans, as these were recognized to have particular
American appeal to Japanese golf customers. The same reasoning was
applied to the sites Ask the Expert section, which contains direct quotes
from US golf professionals regarding golfing tips and Chipshots products.
Recognizing that further market research might indicate a need for translating more of the site, Chipshot wanted to have the option of adapting
these resources in Japanese at a later time.
Ba ck En d

The existing site infrastructure was analyzed to determine what would be

required to achieve full globalization. Chipshot engineers then decided
which alterations they could accommodate without compromising the
quality and functionality of the original English-language site.
Technical

Initially, the most challenging back-end issue was the foreign-language text.
It quickly became clear to Chipshot engineers that correctly displaying
Asian-language text would require experience with character-encoding
issues and double-byte characters. Further, the engineers recognized the
difficulties associated with embedding Japanese characters in HTML, programs, templates, and scripts underlying the existing site. Many sections of
the site required that Japanese text be stored in and recalled from databases.
Achieving full technical functionality and presentational accuracy for the

32 Globalizing an e-Commerce Web Site

Japanese site, it soon became clear, would require Japanese-speaking programmers or an outside vendor for this portion of project.
Managerial

The project offered significant managerial issues, in addition to purely technical challenges. The initial globalization would require coordinating the
work of Chipshots Web developers and site designers with outside translators and editors. Textual changes, such as product-catalog additions and
price updates, were being made to the English site on a regular basis, and
these would need to be reflected in the Japanese site in a timely manner.
Managing globalization would thus be an ongoing effort, in order to ensure
that both native-language and foreign-language sites would always be
synchronized.
Long-Term Development

Chipshot anticipated that changes to the site would involve not only textual
content but also functional features. A functional feature is an element of a
Web site that serves the viewer in an interactive way, often delivering
custom content to a particular visitor to the site. Prominent functional features on the site include:
Shopping System. This feature allows customers to add products to an electronic shopping cart and subsequently purchase all shopping-cart items via
credit-card or customer-account number through a secure connection.
Golf Club Configurator. The Perfect Fit feature allows customers to submit
specific information, such as an individuals height and wrist size, for
ordering custom clubs.
Order Tracking. Customers enter order numbers and track delivery status
of purchases.
Gift Suggestion Center. Customers input information about a person for
whom they wish to purchase a gift and then view a list of appropriate
suggestions.
Account Accessing. Customers review their transaction history and modify
account and billing information.
On [Link], functional features are updated, added, or removed
weekly. Since these modifications would be frequent and ongoing, the
company opted for a maintenance-system approach to globalization. A
maintenance system, as opposed to an ad-hoc nonsystematic approach, formal-

Susan Cheng 33

izes a process of streamlining and integrating all tasks required to globalize a

Web site.
The Web-site classification framework below helps determine the most
appropriate globalization approach. The more interactive and subject to
change a site is, the more appropriate a maintenance system for globalization. The more static and unchanging a site, the more appropriate an adhoc, nonsystematic method.

Interactive

[Link]

Web Apps

E-retailing

E-support

[Link]
News

Ad Campaign
Static

Infrequent Changes

Frequent Changes

In summary, [Link] required a Web-site globalization system that

would: 1) consistently deliver quality, up-to-date foreign-language versions
of their Web site; and 2) flexibly mesh with their existing Web-site development and management process, even as changes to this process were
being made.
O P T I O N S

Chipshot considered the following in-house and outside vendor options for
globalization:
Machine Translation. The first option was to combine readily available
machine-translation and translation-memory tools with in-house Webdevelopment resources. This would mean devoting at least one full-time inhouse developer to effectively applying machine translation, and one fulltime translation manager to supervise ongoing translation. While machine
translation was potentially cost-effective, the company felt it could not rely
on such language-automation tools alone to deliver a foreign-language site.

34 Globalizing an e-Commerce Web Site

Translation Service Bureau. Another option was to combine in-house development resources with outside human translation services. Outsourcing
only the translation necessary for the project would mean devoting significant in-house resources to manage the project and aid with
implementation. Chipshot considered this option unfeasible given the companys limited in-house resources and lack of experience with a project of
this nature.
Software-Localization Firm. Chipshot considered contracting with softwarelocalization companies. However, the company found it difficult to locate a
traditional localization house that could guarantee on-time delivery of a
Japanese e-commerce site. The localization companies generally proposed
to globalize the site by rebuilding it from scratch, and to incur substantial
initial and maintenance costs in the process. For similar reasons, Chipshot
ruled out Web-design and development companies.
Web-Site Globalization Company. Chipshot ultimately chose to work with
Idiom Technologies, a company that offers technology and service products
that enable delivery of end-to-end, integrated Web-site globalization.
Globalization Process

Idiom Technologies globalization approach involved using WorldServer, a

software-based system that facilitates the three main aspects of globalization:
translation, workflow, and delivery. Globalization requires that: 1) site content is presented to translators in a way that allows for high-quality, incontext translation; 2) ongoing translation, editing, and validation of content are efficiently managed; and 3) translated content is properly published
within a Web context and is delivered on time.
Chipshot selected WorldServer for two main benefits: reduced cost and
quick time to implementation. WorldServer eliminates the need to build
and manage a foreign-language site separately from the native-language site
(i.e., it does not incur the cost of creating and maintaining parallel sites).
WorldServer also minimizes the management issues typically involved in
Web globalization through the use of integrated and automated workflow.
As a result, management of one or many versions of a single site remains
centralized and only one Web manager is required for initial and ongoing
management of full-scale globalization.
Based on XML technology, WorldServer facilitates the extraction, distribution, and final synthesis of Web content during translation. WorldServer
implementation involved leveraging Chipshots existing Web-site infrastruc-

Susan Cheng 35

ture while separating globalization issues from the companys regular Web
development.
Mod el Arch it e c tu re

The WorldServer system includes three main components:

WorldTags. These XML tags are added to the original-language Web site at
the HTML level to identify all content in need of translation. These directives are also used to specify content for different target languages and
countries.
WorldManager. This integrated globalization environment is used by translators, content editors, and site managers. A key component of
WorldManager is the Translation Center, the module used by those
involved in translating textual elements. Site managers use WorldManager
to identify content in need of translation and then monitor the distribution
and collection of translation tasks. WorldManager is also used by site managers to review and approve all content for publishing on multiple-language
versions of a global site.
WorldPublisher. WorldTags embedded within the original-language site
speak to this production server that stores translations and works with
existing Web servers to deliver a global Web site.
When a foreign-language user requests a Web page from the original Web
server, WorldTags and WorldPublisher together synthesize a coherent foreign-language version of the page. The new version, delivered in real time,
is comprised of content (i.e., text and graphics) previously translated and
stored in a translation database.
WorldServer simplifies the challenge of supporting double-byte characters
to correctly display Japanese text. The infrastructure software supporting
[Link], like most commonly used Web-site authoring tools, does
not support non-ASCII text on its own. Without WorldServer, all hardcoded strings of text embedded in HTML and scripts would have to be
moved into a resource database before being translated. Without this new
technology, locating Japanese-translated content within the original site
would require extensively rewriting scripts and other code.
WorldServer lets Chipshot sidestep these iterative and QA-intensive processes. Within a unified workflow system, WorldServer can support
multibyte encoding standards including ucs2, utf-8, and other Unicode
standards, and display Asian text using formats such as Shift-JIS, Big5,
and GB.

36 Globalizing an e-Commerce Web Site

Process

Web globalization often involves multiple translators, editors, local-content

authors, developers, and designersideally coordinated by an experienced
Web manager. Globalization has a cyclic workflow that should consist of a
few basic, repeatable actions:
Initial Web Site Review
- content review
- functional review
- code & database review

Identify content on original Web site

for globalization; specify content for
particular target languages/countries

Scan site for content in

need of translation

Assign specific content

elements to translators
Repeat cycle daily,
weekly, or monthly, as
needed
Authorize translation process

Review & validate completed translations

Activate new translations

Initial globalization took approximately three weeks. This was the total time
required to have the site fully functioning and running live as an electronic
retailing site in Japanese. A breakdown of the initial globalization by time
and staff resources is shown below.

Susan Cheng 37

Task

Human Resources

Time
Required

Identify text and graphics in need of

translation

1 project manager
(using WorldTags)

16 hours (2 days)

Scan site and assign identified content to

translators

1 project manager
(using WorldManager)

8 hours (1 day)

Translation and editing

2 translators
(using WorldManager
Translation Center)

128 hours
(16 days: 8 days
per translator)

Review and validation

1 project manager
(using WorldManager)

16 hours (2 days)

Total: 168 hours

(21 days)
Preparing the Site

The project manager first identified translatable items using WorldTags.

WorldTags also mark conditional contentitems that should appear only in
certain language versions of the site. For the initial globalization, Chipshot
wanted only the main sections of the site translated, including the navigation bar and all instructions for online shopping.
WorldTags signal content for globalization
Mark content for
region-specificity

Mark content for

translation

Mark content for

currency conversion

WorldTags also mark date and time information for correct display and formatting for Japan. For currencies, WorldTags marked currency for
automatic conversion into yen, according to a live conversion-rate source.
As WorldTags are HTML-compatible, they were embedded within HTML,
CGI scripts, programs, and other WorldTags.
Scanning for Global Content

After marking the original code with WorldTags, the project manager used
WorldManager to automatically scan the US site for all text and graphics
that had not been previously translated. WorldManager identified items for
translation by crawling the 25-page site in a logical fashion. Site content
was automatically divided into individual translation items, each assigned a
unique identifier. A single item could consist of a single graphic, a short
string of text, a paragraph, or a whole page. During scanning, the project

38 Globalizing an e-Commerce Web Site

manager verified items in need of translation and then used WorldManager

to store the items, preserve their contexts, record any associated comments,
and feed everything to the Translation Center where translators could
access it.
Translation

Two translators were assigned to translate the selected text and graphics and
then edit each others work. WorldManagers Translation Center displays
newly assigned items, now called translation projects, within an organized
interface. Translators were free from code; only text, graphics, and other
pure content resources were shown, allowing them to focus on translation
and not be distracted by HTML. The Translation Center hides HTML and
other programming code from the user without actually separating code
from context, preserving functional integrity.
Translators could view items in the context of the original Web page and
then preview translated items within a new Web page. These features
helped translators perform in-context translation and initiate quality control
from the beginning of the process. The Translation Center also integrated
and tracked all changes and comments made to a translation project.
A Translation Center interface where translators work with text in
context, separately from HTML code

One translator chose to work online within the browser-based Translation

Center interface shown above. Items for translation were separated from
HTML and listed in the lower left, while translations were typed into text
boxes in the lower right. The other translator chose to export and then
import translation, working offline using preferred tools and processes; in
this case the Translation Center was primarily used as a task-management
and quality-control environment.

Susan Cheng 39

Appropriately tagged dates, times, and currencies were converted and formatted by WorldServer and automatically translated according to predefined
formulas. This feature, also known as WorldServer Autoconvert, helped
reduce the total time required to translate site content.
Translation and editing lasted approximately two weeks, including training
in use of the new tool. After translation and editing, the Translation Center
recorded approvals and forwarded translated content to the project manager
for publishing.
Tracking and Validation

The project manager used WorldManager to track the progress of assigned

translation projects and preview translated items in context. Once all translation projects had been completed and approved, the project manager
conducted quality assurance. This involved reviewing and validating completed translations in Web-page context within WorldManager, from a
browser pointed to the testing server.
Activating Translations

Finally, the project manager activated all validated portions of the newly
translated content for publishing on the live global site. After the completed
Japanese version was tested on the live server, it was connected to the
English site via hyperlink. Below are the results of initial implementation.
A screen shot taken from the US-English [Link] home page after
initial globalization

40 Globalizing an e-Commerce Web Site

A screen shot taken from the Japanese [Link] home page after
initial globalization

WorldManager allowed selection and control of basic content differences

between the two versions. For example, site managers could specify that the
Fathers-Day promotion available to US-based customers be published only
on the English site. A Japan-specific promotion was displayed on the Japanese site in its place. Also, recognizing the appeal of cartoon characters to
Japanese customers, managers intentionally positioned Chipshots cartoon
caddy character more prominently on the Japanese site. A free shipping
promotion, contests, and games available only to US customers were displayed only on the English site.
Managing Ongoing Globalization

As content or functionality is added or updated in the original-language

site, the project manager uses WorldManager to detect and initiate globalization of changes for the Japanese version. The process of scanning for new
or changed content, assigning items for translation, and then activating
newly approved translations can be initiated as frequently as desired.
Since the English site was still the main draw and focus, Chipshot continued to make daily changes to the English site. An aggregation of these
changes initially was propagated to the Japanese site approximately every
two weeks, with the expectation that the revision cycle would be shortened
once the marketing campaign for the Japanese site was under way.
During each synchronization or revision cycle, the project manager used
WorldManager to identify new elements in need of translation into Japanese. As content was added to the original site, WorldManager registered
these changes automatically and immediately prompted this content for
translation the next time the project manager logged on to the system. After
logging in to the Translation Center, translators received these elements as
new assignments. The project manager tracked the status of assignments,

Susan Cheng 41

validated completed assignments, and made translations go live on the Japanese site as appropriate. Here is a breakdown of the first resynchronization
process (involving translation of 2,000 words), by time and staff resource:
Task
Scan site and assign updated content to
translators
Translation and editing

Review and validation

C O S T

Human Resources
1 project manager
(using WorldManager)
1 translator
(using WorldManager
Translation Center)
1 project manager
(using management
module)
Total:

Time Required
1 hour
8 hours

1 hour

10 hours

B E N E F I T S

Chipshots actual return on investment for the companys first globalization

project has yet to be determined, as efforts to market the site in Japan have
only just begun. However, a cost-benefit analysis is possible from an implementation standpoint.
Most other e-commerce sites that have globalized at all have employed a
parallel-sites approach. These companies have built and managed foreignlanguage versions separately from the original native-language site. This
requires manual propagation of changes during revision. Overall, the parallel-sites approach would have cost Chipshot, over a three-year period,
approximately five times the resources spent to implement the maintenance
system used for [Link].
The estimated resource costs for initial and ongoing globalization of [Link], using the parallel-sites approach, was estimated by Chipshot Web
managers and developers during initial evaluation of the globalization
project (see below).
Estimated work log for initial globalization
Task
Reengineer database
Rewrite scripts underlying Web site
Extract text and graphics for translation
Translation and editing
Quality assurance

Human Resources
1 developer & 1 translator
1 developer & 1 translator
1 developer & 1 translator
1 developer & 2 translators
1 developer & 1 translator
Total:

Time Required
15 days
25 days
5 days
20 days
20 days
85 days

42 Globalizing an e-Commerce Web Site

Estimated work log for resynchronization

Task
Identify text and graphics in need of
translation
Translation and editing
Rebuild updated sections of Web site
Quality assurance

Human Resources

Time Required

1 developer & 1 translator

1 day

1 developer & 1 translator

1 developer & 1 translator
1 developer & 1 translator
Total:

3 days
2 days
1 day
7 days

The above estimates outline a parallel-site globalization using in-house staff

throughout, except for translation. These estimates, we believe, are similar
to those for a parallel-site approach that gives complete Web-site globalization responsibility to an outside vendor.
Nex t S t ep s

At the outset, Chipshot decided not to market its Japanese site in earnest
until it had been running live and proved functionally successful over a
three-month trial. At present writing, it has been one month since Chipshot began full-fledged marketing. The company plans to increase the
frequency of revision cycles to keep product lines and site features on the
Japanese site more continuously up-to-date with those on the English site.
The globalization of [Link] embodies the cutting-edge skills needed
in communicating across cultural boundaries in the dynamic environment
of the Web. Language professionals are not only required to prepare idiomatic and accurate translations of text that will appear on the Web; they must
also, under the guidance of a targeted marketing strategy, tailor the message
to fit the medium and purpose. Graphic designers must also provide the
design framework to present this message in a culturally appropriate setting.
[Link] also shows how new tools are continually being developed to
meet the intense language demands presented by the World Wide Web.
The technology solution provided here allowed major cost and time savings, while minimizing technical difficulties inherent when multiple parties
get involved in modifying a Web site.
The principles applied in globalizing this e-commerce site should help others in improving process in their own Web globalization. Given the
dynamism of the Web, one can safely assume that new technologies and
solutions will provide further enhancements in the years to come.

L A N G U A G E M A N A G E M E N T
C A S E S T U DY

Section II: Language Management

What Price Globalization?

Managing Costs at Microsoft
David Brooks
Microsoft Corporation

O V E R V I E W

In fiscal 1998, over 60 percent of Microsofts revenues came from markets

outside of the United States. The majority of these revenues come from
non-English-speaking markets, and a key component of Microsofts international strategy has been to lead the industry in the delivery of localized
products to these markets. In that same period, Microsofts revenue from
localized products exceeded US$5 billion. A mere five years ago, these figures were a fraction of what they are today, and as revenues have grown, so
has Microsofts investment in localization.
Localized Product Revenues at Microsoft (in Billions USD)
$6
$5
All Others

$4

Eastern Europe
$3

Latin America
Asia

$2

Western Europe
$1
$0
94

95

96

97

98

99

As Microsofts product range and scope of localization grew, executives

cringed at the rapidly rising cost of localization. Chairman Bill Gates characterized localization as just a linguistic process and expressed frustration
over what appeared to him to be runaway cost increases. Worse yet,
Microsofts ability to ship localized products was being constrained: the
company was having trouble delivering the breadth of localized products
the market wanted and meeting customers demands for prompt, simultaneous release of localized products.

44 What Price Globalization? Managing Costs at Microsoft

T H E

P R O B L E M

When Gates and Paul Maritz, then head of Microsofts Platforms and
Applications product-development group, began to look into localization in
the early 90s, the situation was as follows:

The companys ability to ship localized products was hampered by

delays, bottlenecks inside and outside the company, technical barriers,
and costs.

More products and languages, combined with production bottlenecks,

resulted in projects being prioritized into tiers, with the last tier delayed
up to a year or more after the US product release.

The Redmond, Washington product groups had nominal responsibility

for localizing their products, but the problem fell mainly on the shoulders of the Microsoft teams in Ireland and Asia.

With a few exceptions for secondary languages or products, most

software was localized by Microsoft staff and outsourcing was limited to
manuals and help files.

Despite the lack of metrics, management believed the companys

localization process was inefficient, expensive, and unable to meet
market demands.

Under the sponsorship of Maritz, a campaign (or jihad, in Microsoft-speak)

was launched to bring these problems under control. Maritz asked me to
drive this initiative. I had joined his organization several years earlier as
director of business operations, and had a background in management consulting and cost management. Needless to say, I faced a steep learning curve
in grappling with Microsofts localization problems.
My first step was to find knowledgeable people in the companys localization community. This led to a localization summit in the summer of
1994, which brought together senior localization people from around the
company for two days of discussion.
Early in the game it became clear that localization is not really about translationin fact, it is an extension of product development. Most people
think translation when they think about localization. In a perfect world,
localization would be, as Gates characterized it, just a linguistic process,
but the US-centric design of the US product, unpredictable changes in
schedule, and logistical problems make localization a painful reengineering
process. The chart below summarizes localization spending by discipline at
the start of the jihad. Translation accounted for about a third of total spend-

David Brooks 45

ing, and there was a great deal of skepticism in Redmond regarding the
remainder.
Globalization Spending at Microsoft by Discipline
17%

20%
Engineering
Testing
Translation
Project Management

35%

28%

The summit resulted in a meeting with Gates later that year where a game
plan for attacking localization costs was outlined. The plan was based on the
following findings from the summer meeting:

Localization is not just translationin fact, the translation component

is the more straightforward part of the job.

The majority of localization effort goes into remedial engineering to

enable the US product to work with foreign languages, and testing to
assure nothing breaks during localization.

The logistical aspects of localizationkeeping track of individual files,

managing changes, creating the golden masterrequire extensive
project management.

Changes in the development schedule of the US product wreak havoc

on efforts to plan and execute localization projects.

Localization costs are hard to capture; managing costs was not a

priority, and there was little agreement on how to define or measure
efficiency.

Ma ri tzs Localization Dire ctive

Maritz focused on the engineering aspects of localization, and the goals he

set for the development groups reflected his assessment that US-centric
design was the heart of the problem. He distilled this into a directive to the
product-development teams:

Globalize the US product so reengineering is not required during

localization.

Adopt a common set of localization tools and processes across the

company.

46 What Price Globalization? Managing Costs at Microsoft

Reduce the level of technical skills required to localize products so that

external vendors can execute many localizations in parallel.

Maintain a single codebase across all localized versions of a single

product to assure compatibility across languages.

We knew that the hardest task would be breaking through the US-centric
mentality of the development community. The US is the most competitive
and innovative software market in the world, and software engineers looked
to the US trade press and competitive analyses as a report card on their
work. Globalization and localizability were not major issues in the US, and
in the minds of most developers, these chores were simply a distraction
from the challenges of building new, compelling products. Nonetheless, we
knew that achieving our goals required bandwidth from the development
community dedicated to these issues.
Global Design

Prior to the jihad kickoff, localization was handled as an afterthought by

downstream localization teams and vendors. Developers and engineers in
the core (i.e., US) teams paid little attention to international issues, and
their nave view was that localization equals translation. They were generally unaware of the biases and engineering limitations inherent in the US
product they were building.
One example of engineering limitations is codepage support: all characters,
numerals, punctuation marks, and other symbols commonly used in English
(and many other Western European languages) fit on a single list of 256
characters (the codepage), each of which is addressed with one byte. Japanese uses thousands of symbols, thus the Japanese codepage numbers in the
thousands and requires two bytes to enumerate each character. Thus, the
protocol for addressing specific characters on the codepage must be a twobyte system (also referred to as double-byte encoding).
Codepage issues are pervasive, and double-byte enabling comprised the
bulk of the Japanese localization teams effort. Because so much source code
required modification, extensive retesting was mandatory. As a result, Japanese products were shipped a year or more after the English release, and
inconsistencies between the English and Japanese versions of the same product made it difficult to share files between Japanese and English users. Furthermore, there was very little available time and too few resources to add
features and functions urgently needed by Japanese customers.
Maritzs directive meant building double-byte enabling into the US code, as
well as adopting a variety of other practices to eliminate other engineering
limitations that required reengineering prior to translation. Although this
imposed a burden on US-based developers, it was obvious to Maritz that

David Brooks 47

fixing the problem at the source would be a far more cost-effective and
timely solution.
This initiative went under the heading of globalization, and encompassed all
aspects of the development cycle, including tasks such as product-setup
localization, build process, and test scenarios. Globalization was the cornerstone of the strategy to control costs, and to enable Microsoft to localize
more products without a commensurate increase in staffing or costs.
L o c al i z abi l it y

The problem proved more complex than we first thought. Even a well-globalized product can be difficult to localize: globalization only assures that a
program can accommodate foreign-language datait doesnt assure, for
example, that nothing will break when the user-interface is translated. The
majority of these problems are simple, stupid mistakes that should not
occur in the first place, but that are almost unavoidable.
For example, a well-globalized product may fail when English commands
are replaced with translated versions, because the number of spaces set aside
in the code for the command (i.e., the size of the string buffer) may be
insufficient to handle the translated command. These problems, which we
record and report as bugs, require diagnosis and engineering to identify and
solve. It is easy to automatically adjust the size of the string buffer to the size
of the text, but hundreds of people may work on a piece of Microsoft software and it is inevitable that someone will forget to do this.
Another frustrating localizability problem is related to the use of hot keys.
In many Microsoft products, users can change a text to bold by highlighting
it and simultaneously pressing the Alt and the B keys. This makes intuitive sense in English since bold begins with B, but this is not true for
other languages. Localizing hot keys like this can be difficult: what do you
do when there are several hot-key functions that begin with the same letterchange the name of one of the functions (and, of course, the help files,
manual, and all other places where the function is referenced), use an arbitrary key with no mnemonic value and force the user to learn another arbitrary command, or simply eliminate the hot-key function?
Recognizing this class of problems was embarrassing, but resulted in development of tools and procedures to eliminate the problem at the source
i.e., in the core code. Core teams were asked to take direct ownership for
German and Japanese localization in order to ferret out localization problems, and developers around the company began building tools to find and
eliminate localizability bugs by automatically translating user-interface (UI)
elements into a worst-case nonsense language before shipment of the US
product.

48 What Price Globalization? Managing Costs at Microsoft

L o c al i z at ion P la nn ing

The primary objective of localization is to meet consumer needs in international markets. Decisions about which products to localize, how extensively
to localize them, and the delivery timeframe are driven by market considerations. We define market linguistically rather than geographically: the Spanish market is the aggregate of Spain, Mexico, Argentina, etc.a total of 22
countriesnot just Iberian Spain; German is similarly an aggregate of
Germany, Austria, and part of Switzerland.
The primary criterion for assigning markets to categories is revenue from
localized products. While there is an element of circular logic in using
localized-product revenue as a basis for planning, our experience shows this
is the best single indicator of sales potential. Other parameters, such as population, have little bearing on the software market, but to assess international markets we do look at the number of PCs sold into certain markets,
growth rates in PC sales, and nonquantitative factors such as the degree of
protection offered for intellectual property rights.
We use four major categories (tiers) in the planning process:

Tier 1. The largest international markets (Japanese, German, French)

for which the majority of products are localized.

Tier 2. Markets large enough to justify substantial investment in

localization (Dutch, Korean, Brazilian Portuguese) but an order of
magnitude smaller than Tier 1.

Tier 3. Small but growing markets (Portuguese, Arabic, Hungarian) for

which a subset of products with broad appeal is localized.

Tier 4. Emerging markets of limited potential (Thai, Romanian,

Vietnamese) for which only core products such as Windows are
localized.

For planning purposes we group our products into categories that correspond to market segmentsdesktop products, such as Windows, Office,
and Internet Explorer used by individuals for personal productivity; business systems, such as NT Server and Exchange used by large organizations
to run their operations; development tools used by engineers and developers to create software, etc.
Market tiers and product categories are arranged in a grid; based on the
requirements of each tier and the product characteristics, we decide on the
degree or level of localization. Localization levels, beginning with the lowest, are summarized below (each level is incremental, i.e., the localized
product is both enabled and localized):

David Brooks 49

Enabled. Users can compose documents in their own language, but the
software user-interface and documentation remain in English.

Localized. The user-interface and documentation are translated, but

language-specific tools and content remain in English.

Adapted. The linguistic tools, content, and functions of the software are
revised or re-created for the target market.

For a tier-1 market like France, virtually all desktop products are enabled,
localized, and adapted with French content. Not only are the user-interface
and documentation translated into French, but linguistic, formatting, and
stylistic tools like spell-checkers, business-letter wizards, and Internet links
(in Internet Explorer) are redesigned to conform to and reflect the tastes
and interests of French customers.
A critical dimension of localization planning is the schedule for shipping
localized products. This is usually expressed as the number of days between
the US-product release date and the localized-product ship date. This delay,
referred to as the delta, is critical: products that ship with a small delta
(under 30 days) can ride the wave of publicity surrounding the US product
launch, while products with longer deltas present marketing challenges,
because customers commonly stop buying existing products when new
ones are announced.
Microsoft localizes key products such as Windows and Word into 25 or
more languages, but when the jihad was launched, it was not feasible to execute all projects in parallel. Top priority was assigned to top-tier markets;
emerging markets suffered from long delays (up to a year or more) before
localized products were available.
O pe r a ti o ns

Microsoft handled localization internally until the early 1990s. A single

group in Redmond handled localization for all products, and was staffed
with a small army of young college graduates and expatriates. The International Product Group (IPG) was far from the mainstream and had little
influence on product development. Senior managers and engineers in the
core teams were largely unaware of localization, how much it cost, or how
it was done. Fortunately, the number of products and languages was small,
and international customers were far less demanding than today.
As the scope of localization grew, managers of the product groups became
concerned about growth in IPG headcount. Mike Maples, head of the
product groups before Maritz, anticipated that the amount of localization
work was going to grow rapidly and could not continue in over-the-wall
mode. Maples subsequently split up IPG, divided its headcount among the
product groups, assigned them responsibility for localizing their own prod-

50 What Price Globalization? Managing Costs at Microsoft

ucts, and encouraged outsourcing tasks such as translation. He also established a small central group to support the product groups localization
efforts by finding suitable translation vendors, negotiating master contracts
with the vendors, establishing consistent terminology, etc.
Microsoft executes over a thousand localization projects each year. (A
project is defined as a product/language combinatione.g., French
Windows 98, Polish Internet Explorer 3.0.) We do not view translation
itself as a core skill for Microsoft, and based on Mapless directive we began
to outsource the majority of translation work. Localization is more than
translation, and even the relatively simple task of translating software can
introduce problems and complexities. As a result, we adopted a mixedsourcing model where localization of textual material like printed documentation and online-help files was virtually all outsourced, while software
localization remained in-house. Most other software-related activities such
as testing, compilation, build, and setup localization also remained in-house.
To handle software localization and oversee our European localization vendors, we set up a dedicated localization operation in Ireland in 1988. The
decision to locate in Ireland was based on Irelands physical proximity to
continental Europe, the availability of a well-educated English-speaking
workforce, a good telecommunications infrastructure, tax incentives, and
Irelands pro-business attitude. The Irish team quickly emerged as
Microsofts localization experts, and became highly skilled in fixing the
many engineering glitches and localization bugs inherent in the US product, developing localization tools, and managing vendors of varying sophistication and skill. A similar process later occurred in Japan and elsewhere
in Asia.
Th e Loca li zati on Indu str y

The localization-vendor industry is an essential player in the delivery of

localized products. Over the past 10 years, for-hire localization has become
a US$2 billion industry, and a vital link between software publishers and
their international customers. Microsoft is the largest single customer for
localization services.
In the early 90s the global localization community was a network of
mom-n-pop shops scattered across the globe. Most were tiny, undercapitalized businesses with only a handful of permanent employees. Most work
was done by freelance translators hired on a per-project basis, and the level
of engineering skills was very uneven. Many of these companies started life
as translation houses, and while their business skills were usually good, their
software skills were often inadequate to deal with the thornier problems
challenging Microsofts localization teams.
Other firms, especially in Eastern Europe, were the opposite of the first
group: highly capable software engineers utterly lacking in business skills. I

David Brooks 51

still remember a visit to one firm in the former Soviet bloc. The entire staff,
including many Ph.D. engineers who were refugees from the Russian space
program, was jammed into a single residential apartment. Their local-area
network server was located literally in the toilet, and nobody working there
had ever heard of an income statement.
Our ability to quickly deliver localized products to dozens of markets
depends on effective utilization of the resources in the vendor base. This, in
turn, requires reducing engineering and technical obstacles inherent in the
product, and on the vendors ability to supply qualified staff with the right
mix of linguistic, engineering, and project-management skills.
Getting a Grip

Gates is fond of quoting Andrew Grove, former chairman of Intel: If you

cant measure it, you cant manage it. This was certainly the case with
localization. Prior to the jihad, the little available localization-cost data were
in the form of accounting records compiled at an aggregate level on a fiscalyear basis. These data failed to show how the money was being spent (e.g.,
how much on testing, how much on translation), neglected to address the
project nature of the work, and offered no insight on the impact of increasing complexity of the companys products, shorter deltas, addition of new
languages, etc.
The first step in grappling with the problem was assembling reliable, consistent cost information. Because of the mixed-sourcing model, this meant
capturing internal and external costs. To get beneath the macro view, it was
also necessary to define cost categories that could be used across a wide
variety of products ranging from operating systems like Windows 95 to
games like Monster Truck Madness.
External costs were recorded under a variety of accounts in the general ledger, and a paper trail of invoices was also available. Getting a handle on the
internal costs proved much more difficult: although costs for the localization teams were recorded in aggregate at the department level, none were
categorized by task or assigned to projects.
We interviewed the internal teams to establish a basis for allocating internal
costs. The Irish localization teams responded to this initiative with enthusiasm (they had grown tired of having their IQ questioned). They eagerly
provided the quantitative data they believed would show that the problem
really was in Redmond, where the core code was developed, rather than in
Dublin, where the problems surfaced. Although the allocations were
approximations, they would show Maritz and Gates where their localization
dollars were going and where we needed to focus in order to reduce them.

52 What Price Globalization? Managing Costs at Microsoft

Sor t ing Out the Numbers

To help us understand what we were dealing with, we defined a localization

cost framework based on what was being worked on, and what type of skill
was being applied. The result was a matrix of skills and deliverables similar
to the one below:
Framework for Localization Cost Analysis
Globalization

Software
Adaptation

Localization

Doc/Help

Engineering
Test
Translation
Project Mgmt
Total

We defined our terms as follows:

Globalization: remedial engineering the localization teams performed to
make the products functional in the target languages (sometimes abbreviated to G18N).
Adaptation: the process of adding language-specific or market-specific contente.g., spell-checkers, drivers, file converters.
Localization: translating the user-interfaceuser-interface, integrating all files
into a finished product, and managing the product through the release process (abbreviated L10N).
The task categories were high-level aggregations of the many specific tasks
actually required. For example, in addition to the cost for translators themselves, the Translation category also included proofreading, artwork, and
other activities related to the basic task.
Once the raw cost data had been assembled and categorized, we still had to
invent a way of measuring efficiency. We tried and discarded a variety of
approaches, including comparisons of localization costs: a) with revenues;
b) with the cost of localizing the previous version; and c) with competitors
costs as best we could estimate them. None of these indicators was especially useful because market size has little bearing on localization cost, products change enormously from release to release, quality improves with
experience and turnaround time is shortened, and a multitude of other
factors.
Armed with Gatess dictum that localization is just a linguistic process, we
adopted a metric based on a comparison of actual costs with a hypothetical,
pure-translation benchmark. In response to Gatess challenge, we defined
a gauge for localization efficiency based on a hypothetical cost, assuming

David Brooks 53

Microsoft products were so well engineered that nothing more than translation would be required for localization. We compared this benchmark to
the actual costs, calculated the ratio between actual and benchmark costs,
and used this as our measure of efficiency.
Efficiency Metric Pro Forma
Word Count
Software UI
Documentation
Online Help
Total Words

200,000
300,000
500,000
1,000,000

Actual Cost
Globalization
Adaptation
S/W Localization
Doc/Help
Total Cost

Cost Per Word

Benchmark Cost (words x cost/word)
Efficiency (Actual/Benchmark)

$50,000
75,000
500,000
250,000
$875,000

$0.25
$250,000
3.5x

What We Found

Prior to this effort, localization-cost data had not been collected in a comprehensive fashion. As a result, management found it difficult to accept
what the localization teams had been saying about the amount of effort
consumed in testing, building, and simply keeping track of what was going
on. Instead, they wondered about the IQ of the localization teams. Once
the data from our effort were in hand, the depth of the problem became
apparent. As expected, actual costs exceeded the perfect-world benchmark,
but how and why they exceeded the benchmark challenged the conventional wisdom about what was driving costs.

54 What Price Globalization? Managing Costs at Microsoft

Below is an example of the data we compiled.

x-Factor

x-Factor Analysis of Sample Project

Total
Project Management
Translation
Testing
G18N

Adapt

Engineering
S/W Loc

UA

Total

The height of the bars represents the x-factor for each activity. Within the
software-localization category, for example, the tall bar at the back represents the overall x-factor, and each individual bar represents the contribution of Engineering, Testing, etc. The average x-factor for the project as a
whole is a weighted average of the number of words in the software and the
user-assistance (UA) material. Within the software category, for example,
testing is the largest cost element and the largest contributor to the overall
software x-factor.
Translation of manuals and online-help files was the most visible and easily
understood component of localization. As Microsoft products grew in
complexity, the size of the manuals grew. Conventional wisdom held this
was a major factor behind growth of localization costs. With the cost data
from the study, we discovered that, for most products, the cost of translating
manuals was a relatively minor part of the total spending and was relatively
efficient. As shown in the chart above, the actual cost of localizing the manuals turned out to be a low multiple of the benchmark. Taking into consideration the cost of proofreading, frequent changes, preparation of expensive
artwork, and similar costs, we concluded that Microsofts document-localization process could clearly be improved but was not seriously inefficient.
The software, however, was another story altogether.

David Brooks 55

When the benchmark for software localization was calculated and compared with actual software-localization costs, the results were devastating.
Software-localization costs exceeded the benchmark by a factor of 10. Software testing alone cost several times the benchmark, and the same was true
for project management and other tasks. Software translation costs also
exceeded the benchmark because of rework, technical review, and graphics
work included in the category. The data made it clear that software was the
big cost-reduction opportunity (not the ever-expanding manuals and help
files, as had been believed), and substantiated the claims by the Irish teams
that the cost problem was largely attributable to globalization and localizability problems in the US code.
These data were presented to Gates and Maritz early in 1996. The meeting
was not easy, but we brought the heads of the Ireland teams to the meeting
to explain ad nauseam exactly what all the reengineering and testing was
about and why it was required. Maritz had already seen the data and understood the implications, but Gates hadnt. Suffice it to say he was disappointed to learn of the inefficiency of the overall process implied by the
metrics. He was also disappointed by the persistence of the simple, stupid
mistakes that contributed to extensive engineering and testing, in what
should be an essentially linguistic process.
This marked the turning point in the battle to raise awareness within the
development community. At a follow-up meeting with Gates and Maritz,
senior engineers from each product group were invited to describe their
plans for addressing engineering deficiencies. In early 1997, senior engineers were asked to explain the metrics on their latest products and agree
on targets for their next products.
Impact of the Jihad

Microsofts focus on globalization was initially driven by concerns about

cost. When the underlying cost drivers of localization came into focus, the
leaders of the development community recognized that more was at stake
than costs. Microsofts commitment to deliver localized products in dozens
of global markets was threatened by the engineering hurdles that needed to
be overcome and the scarcity of talent capable of doing the job. By attacking the engineering complexity and skill level required for localization,
Microsoft has made considerable progress toward Gatess goal of making
localization just a linguistic process.

56 What Price Globalization? Managing Costs at Microsoft

The data in the table below summarize efficiency gains achieved for a sample of products since the jihad was launched.

x-Factor

x-Factor Improvements

Product C
Product B
Product A
Pre-jihad

Now

For the sample of products we have been tracking, efficiency gains range
between 3050 percent. Other insights include:

Efficiency gains in documentation localization have been modest. The

gains are attributable primarily to increased recycling of texts from
previous versions, application of technologies such as translation
memory, and standardization of authoring tools.

Much of the improvement in software-localization efficiency is

attributable to effective globalization of the US product and
improvements in localizability as a result of localization-sufficiency
testing of the US product.

Testing costs remain high, but are anticipated to decline as the testing
process itself is globalized. Logistics and project management continue
to be significant cost elements.

Sourcing strategies have been revised to optimize utilization of vendor

resources. Outsourcing for its own sake has been abandoned, as we have
found it more cost-effective to localize certain products (such as highend operating systems) on-site. In some cases we have partnered with
vendors to build dedicated capacity to meet our needs.

Globalization has increased our ability to execute localizations in

parallel, allowing deltas to continue to decline. Internet Explorer, for
example, now routinely ships dozens of language versions within a
matter of days following the US release.

David Brooks 57

Despite rebalancing our sourcing model, localization vendors are a critical

part of our localization machine. In many parts of the world the vendor
base is very thin and is still a constraint on our ability to deliver products to
these markets. The industry has undergone a wave of consolidation,
increasing the possibility of sourcing multiple-language projects to a single
vendor and reducing the logistical burden on Microsoft. The days of a single hand-off remain in the future, as the new generation of vendors is still in
the process of assimilating recent acquisitions.
L o o k ing A h ea d: E xp ec t a ti o ns for t he F ut ure

Globalization is now a basic mission of every product group, but the hope
that localization is just a linguistic process is threatened by changes in the
nature of our products:

Content has become a larger element of our products and, as a result,

the adaptation required for localization is increasing. Products like
Encarta (an interactive, multimedia encyclopedia) are an extreme
example of this, while a new generation of products such as MSN,
Expedia, and other online services push localization to new limits.

The proliferation of the Internet has driven new technologies such as

HTML and XML to the forefront. Documents and content authored in
these languages are a mixture of code and text, and in many ways are
reminiscent of the software of five years ago. These technologies are
evolving rapidly and challenge our ability to build localization tools and
processes to support them.

Linguistic technologies such as speech recognition, natural-languagebased search engines, and machine translation will redefine the scope of
localization. These technologies are being commercialized, and will
soon become mainstream utilities. Localizing these technologies really
means recreating them in each target language. The effort required to
create these tools is substantial and there are no shortcuts.

C L O S I N G

T H O U G H T S

The invention of the printing press eliminated the need for scribes, but created a vastly larger market for books and, consequently, countless new jobs
for authors, editors, illustrators, and other specialized workers. By analogy,
eliminating the engineering obstacles to translation of software and documentation will not make localization obsolete. On the contrary, the
improvement in efficiency and reduction in localization cost will stimulate
demand for localized products and create jobs for experts skilled in adapting
linguistic tools and content to meet the tastes of local markets. Although we
have a firm grip on todays localization challenges, we know our task will
only become more complex and challenging in the future.

L A N G U A G E M A N A G E M E N T
C A S E S T U DY

Multilingual Information Management

at Schneider Automation
Cornelia Hofmann
Ovidius Gesellschaft fr Technische Kommunikation GmbH
Thorsten Mehnert
Tanner Documents
M I M :

T H E

C O N C E P T U A L

F R A M E W O R K

Producing product documentation presents unprecedented challenges

today. Corporate management increasingly understands that this part of
product development is on the critical path, and thus directly affects timeto-market. But besides meeting the time requirements, document management has to cope with increasing volumes of information, needs for different formats and platforms, customer-specific information, communication
in many languages, decentralized production environments, and work
under strong cost pressure.
But the real challenges become visible once we broaden our view from documentation to the wider concept it serves: communication. Communication in turn is based on exchanging information, and the product-relevant
information of tomorrow will be something quite different from what we
call documentation today. Some examples:

Documentation integrated into the product. User documentation was a

separate source of information in the past, but it is becoming an ever
more integrated product component. As product-based information, it
appears as online help in software, as a message presented via a productintegrated display, or as a smart wizard that solves a users problem or
interactively guides one through ones work.

Information-intensive products. Product information provides for more

than just a better understanding of the product or easier interaction
with it. Products require an increasing volume of information for
proper performance and functioning. For example, a small pen that
reads text and translates it into another language contains much
linguistic data; a car equipped with a navigation system needs large
amounts of geographic data. In the case of a service, the deliverable
might in fact be made up entirely of information.

60 Multilingual Information Management at Schneider Automation

Convergence of now distinct types of documentation and means of distribution.

Traditionally, marketing, sales, and customer support each had their
own type of documentation and used separate channels for
communicating with the client. The Web has changed this paradigm.
Information is distributed through intranets/extranets, and the content
from different corporate functions has to be well synchronized and
presented in a coherent fashion. David Brooks of Microsoft puts it this
way: The distinction between what is product [...], what is marketing,
and what is technical material is becoming increasingly blurred.1

Mu lti li ngu al In form ati on Ma nage m en t

These challenges will inexorably alter the way multilingual product information is produced in the future. We will have to manage more interdependencies than today, due to the convergence of product and documentation
and of documentation and support, and due to the need for seamless integration of the many internal and external entities participating in production. But time-to-market requirements will make the increasing number of
interdependencies even more difficult to manage, since they require more
concurrent processes, which in turn leads to a higher level of complexity.
To manage this complexity, we must abandon the silo perspective of product development, marketing communication, technical writing, translation,
and product support. In its place, we must think in terms of a coherent
environment where all entities participate in producing product-relevant
information, i.e., information objects (IO). An IO is a collection of information which is identified as a unit. In principle, everything from a single sentence to a large information product like a manual can be modeled as an IO.
Regardless of the level of abstraction, an IO always meets these criteria:

it targets a certain information consumer (e.g., a user of some

equipment);

it is associated with a business entity (e.g., a certain product, product line,

or corporate function);

it has a communicative goal (e.g., it wants to warn the consumer, or helps

her perform a certain procedure);

it carries content, which is information coded in one or more media

(e.g., text, image, or audio) and which is created for a certain locale
(language/country/culture);

1. LocalizationThe Microsoft Perspective, an interview with David Brooks, in LISA

(Localisation Industry Standards Association) Newsletter, 1997.

Cornelia Hofmann & Thorsten Mehnert 61

it might have some publishing restrictions, i.e., created to appear only in

certain information products (e.g., in manuals but not in online help).

As an example, a typical IO in the Schneider Automation documentation

environmenta fragmentis presented in the figure below.

Technological developments, notably the availability of global delivery platforms like the Web, help us make such an environment reality. They even
allow us to have control over the creation, maintenance, distribution (and
even the consumption) of information, as Rose Lockwood describes it
with the so-called InfoCycle.2 The figure below shows an extended version
of the InfoCycle which we use to briefly introduce our concept of Multilingual Information Management.

2. Lockwood, Rose, Are You Ahead of the Curve?, Language International, Vol. 10 No. 2
(1998).

62 Multilingual Information Management at Schneider Automation

Picture the production of multilingual information as a journey: IOs are

born in Info Development, travel their way to Info Consumption, and will at
some point in time be reused/reauthored in Info Update, and reenter the
InfoCyclepossibly many times. Information will not be created for, say, a
manual and its next version, but systematically engineered to serve many purposes, and thus will have a much longer life span than today.
Derived from the marketing goals of a company, each IO can be associated
with the following goal: acquire an information content, x, in a throughput
time, y, moving from Info Development to Info Delivery at minimal cost. MIM
is a corporate process to establish and operate an environment to meet the
goal of each IO. The environment consists of the necessary processes, systems, and organizational structures.
Each individual company certainly needs a different environment to make
MIM a reality, but we believe the following four spheres will always be
involved:

Subprocesses: designing how IOs are manipulated. The operational

sequences set up to create, deliver, consume, or update information.
Subprocesses include, e.g., authoring, translation, publishing,
distribution.

Information Repository (IR): storing, managing, and retrieving IOs. A

logically central (but certainly physically distributed) database

Cornelia Hofmann & Thorsten Mehnert 63

containing IOs. The information objects will be stored in a

nonproprietary format (e.g., SGML, XML) that will allow reuse of IOs
for different purposes and media. The IR will also provide access
control, versioning, and retrieval functionality.

Workflow and project management: planning and scheduling IO life. Tool

support that allows integrated workflow and project management across
all subprocesses. Nonintegrated workflow and project-management
support (e.g., isolated functionality as part of subprocess-specific tools)
will not meet tomorrows time-to-market goals.

Actors: manipulating IOs. An actor can be a system that automatically

transforms IOs (e.g., a converter creating HTML documents from a
number of IOs). It can also be a human transforming IOs with the
appropriate support of a specialized workbench, such as a writer
working with a writers workbench that allows him to configure,
generate, and write information products; an information manager
who uses a workbench to manage and restructure the IR; or a project
manager whose workbench uses functionality of the workflow and
project management sphere.

MIM R ul es

Successfully operating a MIM means allowing IOs to travel the InfoCycle

efficiently. Before looking at how Schneider Automation approaches MIM,
we will identify some general, company-independent rules that should be
followed in creating and managing IOs.
Make IOs reusable. When writing or translating information, reuse existing
IOs as much as possible:

Manage their content in a nonproprietary, media-neutral way so that

you can use IOs for different purposes/media/formats. Proprietary
formatting tags are straitjackets and often mean less efficient
production.

Ensure that their content follows structural rules so that they can easily
be assembled into larger IOs such as manuals, reference guides, and
online help or even provide input for information systems (e.g., those
that support product configuration).

Automate the information production by generating their content

where possible and economically sensible (e.g., by calculating values in
technical specifications instead of hard-coding the numbers in cases
where values change for different products or local measurement units).

64 Multilingual Information Management at Schneider Automation

Create their content so that it can be used for as many business entities
or products you have to describe as possible.

Specify IOs well and produce them accordingly. Efficient information reuse and
automation is possible only if you produce correct information. Otherwise,
you will multiply errors, and the resulting rework will negate the planned
benefits. Some guidelines:

Always know whom an IO should address and what effect its content
should have on the consumer.

Establish documentation standards to guide their creation, and ensure

they are communicated and followed throughout the InfoCycle.

Carefully select the people who create IO content.

Ensure that IOs are properly validated.

Maintain IO value. IO content might be correct the day you created it, but
you must set up processes that guarantee the content will still be correct
tomorrow. Often IOs are not updated, and the need to maintain their value
is underestimated. Therefore:

Assign clear responsibilities for keeping IOs up-to-date. Make

individuals, and not merely departments, responsible and be certain that
they can oversee the entire InfoCycle.

Change IO content only in a well-defined update process.

Keep in close contact with IOs throughout the InfoCycle. Customer

feedback must be captured and integrated. If you fail to determine
customers desired changes in IO content, IOs will reenter the
InfoCycle, often repeatedly, with a content that fails to meet customer
requirements.

Accelerate IO flow. Meeting time-to-market demands requires IOs to travel

quickly:

Automate workflow (e.g., establish a system that knows everyone in IO

production, assigns predefined tasks to them, and automatically routes
assignments).

Use a project-management system that tracks production of each IO

and avoids redundant project management in the different subprocesses.

Appoint an information-logistics manager responsible for the entire

InfoCycle. Setting up responsibilities for information production in the

Cornelia Hofmann & Thorsten Mehnert 65

participating subprocesses like product development, marketing,

technical writing, translation, etc., is insufficient.
Integrate data islands. Although you may have physically separate data collections, ensure they are connected logically and form a single information repository. Hiding IOs in separate data collections in your own or other
subprocesses means other potentially useful IOs cannot be identified/assessed when new ones are created.
Teach everyone to think in terms of the InfoCycle. For the InfoCycle to operate
successfully, everyone participating in IO production needs to consider the
subsequent information-creation or consumption steps. The logistics manager should support this attitude by appropriate communication and motivational measures.
A P P R O A C H I N G

M I M

A T

S C H N E I D E R

The information production at Schneider Automation is based on the MIM

concept, and to date has been partly implemented. We went through three
development stages, as outlined below:
Requirement

Steps Taken Regarding Production

Environment

Provide technical information in more than Introduction of translation-memory tool

one language
Represent the same technical information in Single-source publishing by enhancing
help files and in manuals
desktop-publishing tool already in use
Come as close as possible to simultaneous
publication in four languages
Represent the same technical information
on various media in many languages
Create coherent information at remote
locations

Development of an SGML-based crossmedia publishing system

This case study describes the final two development stages:

Enhancing the Interleaf desktop-publishing tool to create a singlesource publishing tool (which has now been in use for three years).

Developing an SGML-based cross-media publishing system called TCToolbox, implemented in late 1998.

The Transit translation-memory tool (Star) was part of both approaches.

66 Multilingual Information Management at Schneider Automation

Schn ei der s Te ch ni c al C o mmun ic a t ions Tea m

Schneider Automation, a global provider of automation systems, is a Strategic Business Unit of the French Groupe Schneider. Schneider Automation
was created by combining AEG Modicon (Germany/US) and Tlmcanique (France) under the parent organization Groupe Schneider. Today,
Schneider Automation has R&D and production sites in the US, France,
and Germany. R&D projects are carried out by joint international teams,
which results in individual parts of one product often being developed at
different sites.
Technical Communications, such as R&D, is spread over three sites in three
countries. Technical information on individual parts of a product is always
created at the engineering site. In the case of large systems, the corresponding documentation is created at three sites, in three countries, in three languages, by writers separated by thousands of miles.
This particular situation of the Technical Communications Teamnow
commonplace in international businessled to development of the TCToolbox. The first steps were taken under slightly easier circumstances:
fewer output media were required and information was created mainly at a
single site.
Fi r s t A p pro ac h : D o c um en ta t ion a nd H e lp f ro m a Si ngl e S o urc e

The problem. The project that raised the issue of single-source publishing was
the documentation for new programming software under Windows, which
required localization into four languages. We were tasked with creating help
files and manuals, and translating them and the software user-interface into
three languages. A further challenge: the software was created in English
and the first released language version was to be English, while most documentation was to be written in Germany in German.
Three main challenges were clear from the beginning:

We had to publish the first translated version of the documentation

before the source-language version.

We had to create help files and manuals simultaneously, although we

had resources for only one medium.

We had to translate into more languages than beforeand the language

versions had to be published in quick succession.

The solution. We pursued two basic ideas to achieve our goal:

Create help files and manuals from one single source.

Cornelia Hofmann & Thorsten Mehnert 67

Make the translation process as quick and efficient as possible.

Let us examine the realization of both ideas in detail.

Create help files and manuals from a single source.
This idea was realized by developing an enhancement to Interleaf that
enables creation of help files from the Interleaf source. With this enhancement, a writer using Interleaf can key in meta information for help files,
such as topic titles, topic IDs, keywords, etc. While writing a classical manual with chapters and subchapters, the writer can also define a topic in the
help file, define information that will appear in the help file and not the
manual (or vice versa) or both, and also create help hyperlinks (jumps)
beyond the manuals traditional cross-references (see page ). While
editing, the writer can see only one layout on her screen, one that approximates the appearance of the manuals. Layout of the online help, however,
differs completely and is adapted to the common help style.
Not all information would appear in the online help and the manuals:

Software installation instructions are not included in online help, since

the software is already installed when the online help is used.

The detailed menu and dialog descriptions are not part of a manual,
since they are mainly required when the software is used.

With our Interleaf enhancement, the single source contains all information,
including meta information. We gained two advantages with this tool:

Information must be created only once.

Translating one Interleaf file represents translating a help and a manual

file in a single step.

The latter advantage also contributed to achieving our second goal.

Make the translation process as quick and efficient as possible.
This goal was met by applying a translation-memory tool.3 Since Transit
from Star allowed customization to our needs, we introduced this tool and
incorporated it into our documentation process as shown in the figure
below.

3. We will not describe this type of tool here; refer to the LCC KnowledgeBase at
[Link] for detailed information.

68 Multilingual Information Management at Schneider Automation

Interleaf
Creation of
Single
Source

Interleaf Files

Paper

RTF Files

Help

Transit

This resulted in the following advantages:

Translation of a new version of a text used all the translated text of the
previous version. Nothing had to be translated more than once, and
changes did not have to be filtered out. This saved time and money.

Terminology was better maintained, which increased translation

consistency and quality.

Reformatting of translated documents was almost no longer necessary,

which resulted in major time savings.

With Transit, we could create and translate in parallel. Even at times when
many changes were still being made to the software, translating did not represent a waste of time or money, since we knew the exact location of the
changes and could focus any rework on those passages. In this way, we
translated some thousand pages from German into English, while still creating in German. When the software was released, the English manuals and
help files were ready, along with their German counterparts.
We also used Transit to translate the user-interface of the software. The
main advantage was the ability to create a terminology database containing
only menu and dialog terms, which could then be used for translation of all
IOs in the project. The terminology database ensured consistency between
all these IOs, regardless of whether they belonged to the user-interface,
help files, or manuals.
The results. We published the first translation before the source-language
version, and the other language versions shortly thereafter. With our
method, we managed to achieve intervals of only one month between the
language releases, and the first translation (English) was ready simultaneously with the English software version.
We overcame the challenge of scarce resources as well. Although the writers
had to learn to write manuals and help files simultaneously, in the end they

Cornelia Hofmann & Thorsten Mehnert 69

mastered this technique. Of course, creating manual and help files simultaneously took longer than one of those tasks alone. But the time needed was
far shorter than that required for writing one after the other, even taking
into account that large parts of the text can be copied. We could thus complete the project using only existing resources.
Second App roach: Cross-Medi a Pu bl ish ing f rom a Si ngl e S ource

The problem. The impending need to place our technical information on the
Internet, and an increasing number of projects handled at an international
level, led to the decision to convert to an SGML-based cross-media publishing system.
The project we now faced was again documenting new programming software under Windows, with localization into four languages. We were
tasked with creating help files, HTML collections, and manuals, and translating each of these and the software user-interface into three languages.
The main challenges we knew from our first approach were still present,
with several new ones added:

We still had to publish the first translated version of the documentation

(English) before the source-language version. But this time, we had two
source languages, since some parts of the documentation were written
in German, others in French.

We still had to create help files and manuals simultaneously. But this
time, we also had to deliver HTML collections, and we still only had
resources for creating information products in one medium.

Again, the language versions had to be published in quick succession.

The solution. The TC-Toolbox controls the following types of IOs (see the
figures below):

Fragment: a physical file containing information of a certain revision,

version, and language.

Information Unit (IU): the totality of all revisions, versions, and

translations of a single fragment.

Collection: a logical assembly of several IUs.

Information product: an assembly of fragments, selected and laid out for a

certain destination medium (e.g., a manual) and language. From each
IU of a certain collection, exactly one fragment of a defined revision,

70 Multilingual Information Management at Schneider Automation

version, and language is selected to participate in a certain information

product.
Traditional chapters are split into small fragments. This allows several writers (at different sites) to work together on one chapter and reuse the fragments. The fragments are created by the writers independently of the
destination medium. Splitting into fragments is not done at the level of
SGML elements, but at a level that can be compared with the traditional
subchapter level, as illustrated below:
Version 1
Revisions

Version 2

3
Fragments

Translations

Collection
1

2
1

2
1

3
2

2
1

IU5
IU4
IU3
3
IU2
3
IU1
3

Fragments
of IU1 to IU5
for e.g.: Manual,
Version 1,
German

Information product

The core of the TC-Toolbox is the document-management system (DMS).

The DMS controls all IOs and is the starting point for all user actions. An
SGML database is unnecessary since our fragments more or less represent
subchapters.

Cornelia Hofmann & Thorsten Mehnert 71

Generate information
product

Create and check out

DMS

FM + SGML
Check in
Creation of single
source

Pure SGML Files

FM + SGML

Paper

Windows Help Files

Help

HTML files, ready to view

Check out
for translation

Internet

Check in after
translation

Transit

How writers work with the TC-Toolbox. The writers main activities are:

Editing fragments. We use Adobe FrameMaker+SGML as an editor. The

editing application contains a DTD which allows editing all destination
media in parallel. The layout presented to the writer on screen
resembles the layout of the final manual, but is not identical.

Translating fragments. Translation is assisted by the Transit tool (as in our

first approach). Translation memories and terminology databases are
controlled by the DMS.

Building information products. To build an information product, the writer

must create a list of fragments that should be included in the information
product. Creation of information-product files in formats such as
FrameMaker for paper + PDF, RTF for Windows help, and HTML for
Internet or CD-Rom is done by filtering processes or conversion (metamorphosis) into DTDs that are adapted to the medium. In addition, the
FrameMaker application for paper output differs from the editing
application.
Advantages of TC-Toolbox architecture. Separating editing and production
applications provides maximum flexibility in the tools used and the ability
to integrate other output formats into the TC-Toolbox. This separation
also allows optimal adaptation of the resulting information products to the
destination medium.
Field data. The TC-Toolbox was implemented at three sites and used in the
project mentioned above. Our initial experience shows:

Cross-media writing requires advanced training and experience. It

requires a change of thinking achieved only through practice.

Cross-site writing needs a strictly defined style guide. Style naturally

differs from writer to writer even at a single site, and all the more so

72 Multilingual Information Management at Schneider Automation

from culture to culture and language to language. To create information

products in a consistent style is nontrivial, and requires substantial work
beforehand.
The question whether cross-media publishing is worth the investment is
considered in the Quantitative Assessment section below.
Crea t in g Corre ct I nfor ma t ion a nd Ma i nt ai ni ng I t s Va lu e

Reusing information and automating its production can only be done efficiently if: a) correct information is produced; and b) production setup guarantees that this information will still be correct tomorrow. Otherwise,
errors are multiplied, and the resulting rework might offset all the benefits
one hoped to achieve through reuse.
Maintaining the value of information is not a single task, but requires that
the whole InfoCycle be designed to manage information accordingly, as
highlighted below.
Problem-focused information. Since there is no point in creating and managing
information that has little benefit to the client, every IO we create has a
clear definition of: a) the information consumer it must target; and b) the
IOs communicative goal.
This approach is currently being refined and expanded throughout the
entire product lifecycle. We have also observed how customers information
needs vary in different phases (e.g., presales, sales, installation, operation,
and migration to another product). We intend to meet these changing
needs by structuring the information so that each IO will be attributed to a
certain lifecycle phase and will satisfy the customers information need.
Associating IOs to lifecycle phases will allow us to retrieve information
with high accuracy.
Uniform production environment. Schneiders documentation production is
decentralized and must cope with the additional challenge of integrating
people from companies that had formerly been separate entities. Efficient
production requires a uniform environment with standardized processes.
The TC-Toolbox assists in this effort, but in itself is insufficient to synchronize our activities. Therefore we are currently developing a Writers Manual that will ensure consistent treatment of the following at all
production sites:

Writing and translation processes

TC-Toolbox usage

Cornelia Hofmann & Thorsten Mehnert 73

Layout styles

Writing styles

Qualification of content providers. To reuse information efficiently, our product

information needs to be of high quality, the cornerstone of which is clearly
well-qualified talent to produce it. Our skill profile for translators includes:

Technical expertise in automation technology

Demonstrated understanding of our products

Track record as a professional translator

Native speaker of the target language

Familiarity with PCs

Such individuals are rare, and we often must compromise. Lack of expert
knowledge about automation technology must often be compensated for by
close interaction between translator and writer.
Review process. A well-defined and functional review is essential to the MIM
concept. Reviewing source fragments and translations consists of two distinct phases, a technical review and a language review.
The technical review verifies correct content:

The procedures which the user is to follow are correct, understandable

and complete without misleading information.

The specifications are correct and complete.

Although a first review should be done locally with the source file, the
complete international review can only be done by reviewing translations.
Clearly, changes demanded by the technical review must be made in the
source fragment first, and then in the translations.
The language review verifies correctness of the translation itself. Does the
translation express the same content as the source? Has defined terminology
been used? These are the main issues apart from formal checks like spelling
and grammar. One remark concerning the reviewers comments: in general
it may be necessary to translate the comments, since a writer may not be
able to understand the comments in another language. We have not found
this necessary to date, since the writers involved in these projects were
bilingual.

74 Multilingual Information Management at Schneider Automation

Rol es a nd Re s po n s ibi l it i es i n MI M

Our production environment comprises the following roles:

Role

Responsibility

Writer/Project Manager

Create IOs for an information product in one language.

Plan, prepare, and manage translation work of this product
for all other language versions.

Translator

Translate information products and participate in

terminology work. For qualifications, see above.

Reviewer

Check source fragments and translations (reviewers are

mainly engineers and marketers; a language reviewer must
be a native speaker of the target language).

Administrative Support

Select translators, verify their availability for a project, and

handle quotes, orders, and invoices.

Technical MIM Expert

Monitor use of TC-Toolbox and manage its ongoing

development.

Logistics Manager

Plan and design information flow through the InfoCycle

along with representatives from product development,
marketing, support, etc.

Given their key role in our MIM, writers take on additional tasks at
Schneider, primarily in the project-management area:

A writer leads a team to define and maintain terminology for the

systems she controls.

The writer partners with the translator throughout the translation.

She qualifies the translators product knowledge, and is the first contact
for the translator in case of questions.

The writer prepares the fragment translation because she is the only
one who knows when it is ready for translation.

The writer initiates and follows up the review processes.

MI M E vol ut i on

Which MIM rules have we implemented and how? The following table
shows implementation in the first approach, the progress made in the second, and an outlook on TC-Toolbox Version 2.

Cornelia Hofmann & Thorsten Mehnert 75

MIM Rules
(detailed above)

1st Approach
(Interleaf
enhancement)

2nd Approach
(TC-Toolbox V1)

Outlook for TCToolbox V2

Make IOs reusable.

Manage their
content in a nonproprietary, medianeutral way.

One Interleaf source One SGML-source for all media; SGML is

for help files and
nonproprietary. Content and format are
paper, but Interleaf is strictly separated in SGML.
still proprietary.
Content and format
are combined in
Interleaf.

Ensure their content

follows structural
rules.

To a very low
degree.

Create their content

so it can be used for
many business
entities.

For similar Schneider Automation products, the respective

information products are created using attribute control, i.e.,
product names and special features can be hidden according to the
actual contents.

Fragments are created according to

structural rules as implemented by the DTD.
Fragments are smaller units than traditional
subchapters. They are stored in the database
and may be linked to form many different
collections.

Specify IOs well and produce them accordingly.

Always know whom Our documentation is structured according
an IO should address. to the information goal and the target-user
groups.

IOs are assigned

clearly defined
attributes according
to the actions the
user carries out with
the product. This
improves the users
ability to search for
special IOs
supporting individual
activities.

Have documentation A writers manual

A writers manual will be created, as
standards set up to
described the
outlined above.
guide their creation. following:
Creation and
translation processes
Layout styles
Writing style was not
described. The
writers tried to
achieve a common
style via close
communication.
Carefully select those Our writers specialize in Programmable Logical Controller (PLC)
who create IO
technology. The writers first priority, however, is to specialize in
content.
documenting information and using information technology.
We aim at long-term relationships with translators to improve their
knowledge of PLC technology.

76 Multilingual Information Management at Schneider Automation

MIM Rules
(detailed above)
Ensure that IOs are
properly validated.

1st Approach
(Interleaf
enhancement)

2nd Approach
(TC-Toolbox V1)

IOs pass through technical and language

reviews. The review process described above
is not yet controlled by the system.
However, the 2nd approach allows
validation based on the SGML DTD.

Outlook for TCToolbox V2

The review process
itself will be
controlled by the
system to a certain
extent.

Maintain IO value.
Assign clear
responsibilities for
keeping IOs up-todate.

There are clear responsibilities:

for authoring each fragment: writer
for translating each fragment: translator
for collections and information products:
project manager (writer)

Change IO content
only in a welldefined update
process.

Writers work under

a defined update
process, with only
little administration
taken on by the
system.

Responsibility for
creation of
information will also
be given to people
outside Technical
Communications.

Revisions and versions are administrated by

the DMS.

Keep in close contact Reply forms asking for customer feedback

with IOs throughout are attached to the manuals.
the InfoCycle.
Technical Communications stays in close
contact with the Product Support
Department which receives most of the
customer feedback.

Closer integration
with Customer
Support will increase
feedback. Through
use of the Internet,
we expect more and
more direct customer
feedback.

Accelerate IO flow.
Automate workflow.

Parts of the
workflow may be
integrated into TCToolbox.

Cornelia Hofmann & Thorsten Mehnert 77

MIM Rules
(detailed above)

1st Approach
(Interleaf
enhancement)

2nd Approach
(TC-Toolbox V1)

Appoint an
information logistics
manager.

Outlook for TCToolbox V2

Our future
realization will
merge our data
collections with
those of other parts
of the organization,
building a single,
logical technical
information
repository. A logistics
manager will be
given responsibility
for the data
repository and the
complete InfoCycle.

Integrate data islands.

In both approaches, data collections are created and managed fairly independently from
each other.
Make everyone think in terms of the InfoCycle.
One individual from Technical Communications is responsible for a given international
documentation project. This includes responsibility for all IOs involved.
Quantitative Assessment

Brief economic analysis of our cross-media publishing system. Does cross-media

publishing save money? Without it, parallel creation/translation of documents for several media would be necessary, resulting in redundant, avoidable activities. The following tables shows a comparative calculation of the
estimated costs (10,000 pages) per year and for 15 writers:
Per 10,000 Pages
Creation
1 target format
2 target formats
3 target formats
Translation
1 lang., 1 target format
1 lang., 2 target formats
1 lang., 3 target formats
3 lang., 2 target formats
3 lang., 3 target formats
Total (4 lang. incl. source, 2
target formats)
Total (4 lang. incl. source, 3
target formats)

Costs without
Cross-Media
Publishing

Costs with
Cross-Media
Publishing

Savings through
Cross-Media
Publishing

DM 3,000,000
DM 4,000,000
DM 5,000,000

DM 3,000,000
DM 3,300,000
DM 3,600,000

DM 700,000
DM 1,400,000

DM 500,000
DM 900,000
DM 1,300,000
DM 2,700,000
DM 3,900,000

DM 500,000
DM 500,000
DM 500,000
DM 1,500,000
DM 1,500,000

DM 400,000
DM 800,000
DM 1,200,000
DM 2,400,000

DM 6,700,000

DM 4,800,000

DM 1,900,000

DM 8,900,000

DM 5,100,000

DM 3,800,000

78 Multilingual Information Management at Schneider Automation

Within one year, cross-media publishing for two destination formats will
save DM 1.9 million based on a conservative calculation. The calculation is
based on the assumption that the creation costs with cross-media publishing
for two destination formats is 10 percent higher compared to one destination format. The costs for the case without cross-media publishing, however, have been calculated at a very low level, since updating the content to
a new product version in two destination formats very often costs nearly
twice as much as updating similar information in just one destination format. During initial creation of information, the cost difference is not as dramatic, but we experience updates more often than first-time creation.
Quantitative assessments of our approaches. The following table shows the cost
and time savings of our two approaches, along with expected savings of the
next MIM development phase. These aspects characterize TC-Toolbox V2:

Better IO reuse

Higher integration in managing translation memories and terminology

Intuitive user-interface offers better support for writers during editing

1st Approach
(Interleaf
enhancement)
cost savings

DM 1.9 million in one

year

2nd Approach
(TC-Toolbox V1)

In preparation
(TC-Toolbox V2)

> DM 3 million in
one year

> DM 3 million in
one year

costs of system DM 0.2 million

development &
implementation

DM 0.9 million

ca. DM 0.4 million

costs of
conversion of
legacy data

DM 0.5 million

DM 1.1 million

none

time saving

20% of creation time

45% of translation time

30% of creation time 40% of creation time

60% of translation time (est.)
65% of translation
time (est.)

Cornelia Hofmann & Thorsten Mehnert 79

Glossar y
Term

Definition

Collection

A logical assembly of a number of IUs.

Cross-Media Publishing

Publishing information for several media from a single

source.

Document Management
System (DMS)

A DMS organizes and manages IOs.

Document Type Description

(DTD)

Expression of SGML-based structural and semantic rules

which IOs must obey.

Fragment

A physical file containing information of a certain

revision, version, and language.

FrameMaker+SGML

Desktop-publishing tool by Adobe.

Information Object (IO)

A collection of information identified as a unit. In

principle, everything from a single sentence to a large
information product like a manual can be modeled as an
IO. Typical IOs at Schneider Automation are fragments,
information units, collections, and information products.

Information Product

An assembly of fragments, selected and laid out for a

certain destination medium (e.g., a manual), version, and
language.

Information Unit (IU)

The totality of all revisions, versions, and translations of

one fragment.

Interleaf

Desktop-publishing tool by Interleaf Inc.

PLC

Programmable Logical Controllers.

Revision

A fragment in a draft or intermediate status. A revision

will not be published.

Standard Generalized Markup

Language (SGML)

The international standard (ISO 8879) defines SGML as

a language for document representation that formalizes
markup and frees it of system and processing
dependencies.

TC-Toolbox

SGML-based cross-media publishing system, owned by

Schneider Automation, developed by Ovidius GmbH,
Berlin.

Translation Memory System

Focuses on increasing the translators productivity and

typically offers features for reuse of already produced
translations, preservation of a documents format during
translation, automatic search for terminology, etc.
Synonym: translator workbench.

Version

A fragment ready for publishing. Fragments will only be

published as part of an information product.

L A N G U A G E M A N A G E M E N T
C A S E S T U DY

The Role of Translation in

an International Organization
Andrew Joscelyne
Equipe Consortium Ltd.

In a rapidly changing, globalizing world, you might think that the translation departments in the worlds international organizations would adopt the
same cost-cutting and technology-driven solutions as their private-sector
counterparts.
You would be partly right. The European Commissions Service de Traduction
is a prime example. They recently deployed an information-gisting service, which provides machine translation (Systran-based) for Commission
staff who need rapid raw translation delivered over an in-house network. It
is called gisting, since the computer-generated translation generally provides enough substance that the reader gets the gist of what is being said.
It is also no secret that in 1998, a team from the United Nations translation
department commissioned a report on the use and viability of the machine
and machine-assisted translation service at the EC, presumably to examine
the applicability of such technology to its own needs.
Similarly, calls for tenders are regularly issued by large international organizations for external translation suppliers. They often require assistance in
specific language combinations, or in machine-translation post-editing, to
offset pressure on in-house teams. In line with the general trend, translation
volume has grown exponentially across the spectrum of international organizations.
Despite these shared features of the translation landscape, there are specific
aspects of language-transfer management in international organizations that
are not echoed in private-sector practice. This is largely because of the role
translation plays in these organizations, a translation ethos that quite often
differs from market-driven private-sector translation departments.

82 The Role of Translation in an International Organization

By focusing on the Organization for Economic Cooperation and Development (OECD),1 this case study attempts to show that the specific nature of
language, and a fortiori translation, in these organizations determines the
options available for translation management.
A M ul t il i ng ua l M an dat e

Most larger and older international organizations differ from corporate

users of language services in one way: built into their founding charters is a
clause enumerating the organizations official languages. The United
Nations, for example, has six working languages; the World Bank, the IMF,
NATO, and OECD each work in two official languages. We call this property foundational multilinguality.
Many global organizations (as opposed to regional international organizations in Africa, Asia, and Latin America) were created in the aftermath of
World War II, following the creation of the United Nations and the 1944
Bretton Woods conference. Their foundational multilinguality tends to
reflect the linguistic configuration of diplomacy at that time, with French
and English chosen as the major working languages, rather than the German or Japanese of the recently vanquished nations.
This is not the place to analyze the historical and political sources of language choice in such international organizations, or to survey the range of
possible forms of foundational multilinguality (or, possibly its replacement
by de facto rather than de jure multilinguality in more recently founded organizations). For the purposes of this study, we shall assume that foundational
multilinguality is valid for the specific case of language management we
shall examinethat of the OECD.
Foundational multilinguality as we define it is rarely, if ever, found in the
bylaws of a private company, although many multinational corporations
encourage multilingual practice by staff in order to pursue their business
goals. Since corporate ownership, geographies, market range, and business
lines are in near-permanent flux, any decision to stipulate linguistic practice
as constitutive of a corporation would be tantamount to business suicide.

1. Ren Prioux, Head of the Translation Division at the OECD in Paris, provided me with
a comprehensive tutorial about translation at the OECD and international organizations in
general. His willingness to share and explain these procedures allowed me to write this account. Any shortcomings and inaccuracies are my own.

Andrew Joscelyne 83
Tr ans l a ti o n P rese r ve s

The continued identity and existence of international organizations, however, is largely predicated on balancing the interests of nation-state members. This means that explicit foundational multilingualitywhat is written
into the organizations founding charterwill rarely be questioned or modified in response to market or other external pressures. Any translation practice set up under such foundational multilinguality therefore tends to
endure, whatever the changes in the external balance of linguistic power.
A possible threat to the linguistic status quo of international organizations is
a voted decision at the UN allowing Japan and Germany to sit on the Security Council, due to their clout on the world economic stage. If accepted,
the language range of the UN would have to expand to accept German and
Japanese, and possibly European and Brazilian Portuguese.
Similarly, some may claim that the diplomatic, political, and economic role
of the French language has declined since 1945; hence there would be no
overriding statistical or strategic reason for French to be a consistent core
member of the family of foundational languages.
The evidence so far suggests, however, that foundational multilinguality is
highly resistant if not immune to such changes. As a result, the translation
departments of these organizations form a unique translation preserve
which operates with significant autonomy compared with the rest of the
worlds translation activities.
I N S I D E

T H E

O E C D

The OECD has been described as a rich-mans club of like-minded memberspart think-tank, part monitoring agency, and part nonacademic university. It groups 29 of the worlds most developed countries in an
organization that provides governments with a platform to discuss, develop,
and refine economic and social policy.
Founded in 1961, it replaced the Organization for European Economic
Cooperation (OEEC), which was formed to administer US and Canadian
aid under the Marshall Plan for the reconstruction of Europe after World
War II. The member countries each send an ambassador to the OECD, and
each contributes to the ca. US$200 million annual budget.
The OECDs purpose is to build strong economic policies in its member
countries, improve efficiency, hone market systems, expand free trade, and
contribute to the development of industrialized and developing countries.
It functions through a cascade of committees, the preeminent one being the

84 The Role of Translation in an International Organization

Council, which has decision-making power and works at the ministerial

level with member countries. Some 200 other specialist committees, working groups, and expert groups deal with specific issues of policy, with each
producing a steady stream of knowledge in the form of documents.
The OECD is run by the Paris-based Secretariat, which fields 1,850 staff
members to directly or indirectly support committee activities. Seven hundred economists, scientists, lawyers, and other professional staff, mainly
based in a dozen major directorates, provide research and analysis. An additional six general secretariat units and a half-dozen semi-autonomous bodies produce knowledge eventually translated by the Translation Division, as
we shall refer to it.
Kn owle dge: Instrum en ta l v s. Di scu rsive

Knowledge production is used here as a catch-all to cover the OECDs core

activities. However, it has sufficient resonance to capture a significant feature of such organizations, and shed light on how they differ fundamentally
from many private-sector organizations.
Today, knowledge management is a key operational feature of any organization: companies, governments, associations, even families, produce knowledge of one sort or another, or can be engineered to produce such
knowledge by configuring information-technology systems to capture it.
Information from your credit-card transactions can be transformed into
knowledge about you as a desiring, buying, preferring customer by a computer system. The result, instrumental knowledge, serves some further purpose
of the knowledge producer.
At the OECD, however, the production of knowledge (almost totally
embodied in documents) is the organizations end purpose, and it takes a
discursive form, packaged for human (rather than, for example, machine)
understanding. In this sense it operates like a consulting or market-research
firm rather than a manufacturing company. Like any other organization, it
will inevitably publish knowledge about itself (instrumental knowledge) to
achieve some other end, though its primary task is to produce knowledge as
the end product of its own processes.
This knowledge is delivered to its clients in the form of reports on areas of
concern, sets of statistics, and reasoned conclusions about the relative significance of this quantitative and qualitative information. Thus, the OECD is,
for our purposes, a discursive knowledge-producing organization.
Where discursive knowledge is embodied in figures, words, statistics, and
reports, the organizations knowledge-producing function intersects with its

Andrew Joscelyne 85

foundational multilinguality. At the point of intersection stands the Translation Division.

Tr ansl a ti on at Ce nte r-Stage

In a very real sense, the Translation Division stands at the center of the
organizations activities, acting as a quality filter through which the organizations bilingual discourse passes. Four key aspects of knowledge production at OECD effectively empower translators and endow them with a
status they rarely enjoy in private-sector work:

L2 Authoring.2 Many authors of the source material, i.e., the knowledge

producers, write in a non-native languagemainly English, one of the
foundational languages of the organization. In other words, the
linguistic expression of the knowledge produced is often nonstandard.

Committee-Authored Documents. This does not refer simply to multipleauthor documents, which are increasingly frequent in cooperative work
environments. Rather, it refers to documents which have been serially,
rather than synchronously, edited, often by hand, by a cascade of
different committees, each with their red-penning habits and linguistic
(often L2-authoring) quirks.

Multiculturalism. The organization produces knowledge based on

material that is itself the product of other traditions of knowledge
production, in particular knowledge produced by client members of
the organization. As the OECD has 29 members, its products allude to,
draw on, and exploit extremely culturally heterogeneous information.
Unlike companies that might have a single corporate idiom
disseminated to end-readers in a broad array of languages, the OECD
must occasionally incorporate a broad array of idioms into a document
which is then translated into just one other language.

High Referentiality Index. This is a typical property of the way

knowledge is distributed within documents. Since part of the
knowledge produced by the OECD involves legally binding
resolutions, international conventions, and the like, it inevitably refers
to previous OECD documents and those of other international
organizations.

2. Ren Prioux uses the French term allophonie. However, an English borrowing might cause
confusion with the linguistic term allophone, which refers to the phonetic variants of a single phoneme.

86 The Role of Translation in an International Organization

The Translation Division thus does not simply translate all relevant documents into its two official languages, but serves as a quality-assurance filter
in delivering the organizations major products.
Tr ans l a ti o n- D i vi s io n S t r uc t ure

The OECDs Translation Division was founded concurrently with the

organization in 1961. In 1996 it was integrated into a more general Language and Conference Services Division. The Conference Services department takes charge of all interpretation requirements within the
organization, but there is no direct operational association between the two
departments.
The Translation Division is expected to supply specific linguistic services to
any of the 20 or so in-house units (often called Directorates) which cover
the following subject areas: food, agriculture and fisheries, money laundering, social policy, public management, science, technology and industry, statistics, regions, cities and countryside, relations with nonmembers,
economics and long-term analysis, sustainable development, energy (the
OECD includes the International Energy Agency and the Nuclear Energy
Agency), environment, finance, investment, taxation and competition, regulatory reform, and electronic commerce.
The Translation Division consists of four language sections and a relatively
well-staffed Reference and Terminology Unit (RTU), which provides
comprehensive translation support. The workforce breaks down as follows:
Section

Staff Members

French

53

English

11

German

13

Reference/Terminology

Management

At the OECD, and at most other international organizations, translators are

recruited as knowledge workers with a similar civil-service ranking within
the hierarchy to that of engineers and lawyers. They are viewed as skilled
professionals rather than subordinates to the more important work of
knowledge producers.

Andrew Joscelyne 87

The translators act as an intermediary in the knowledge value chain, as the

following diagram shows. The two-way arrows indicate that the translator
will be involved in various cognitively rich feedback loopschecking with
the authors, inventing neologisms as new concepts emerge,3 establishing
official usage, and reusing finalized knowledge products to decode new
knowledge:

Author

Translator

Knowledge
Product

Consider, as a point of contrast, the translators role in an organization with

no foundational multilinguality. In such organizations, translation will be
subordinate to rather than coordinate with the knowledge-producing chain:
documents will by and large be finalized before translation, and language
transfer may be relatively mechanical, cognitively independent of the
knowledge-production chain:
Author

Knowledge
Product

Translator

Recruitment. To recruit the best staff, the Translation Division uses a fairly
elaborate and time-consuming vetting procedure. Candidates must pass
written and oral examinations, given in a number of separate centers
(including Washington, DC) where the organization has satellite offices.
The 1998 recruitment drive attracted 400 candidates, of whom eight were
shortlisted and four finally chosen. This is fairly standard for international
organizations, where about one percent of candidates are hired.
The Translation Division looks for people able to analyze a document and
convey the information clearly and accurately. The existence of L2, committee-driven authorship means that source documents must first be understood, and then translated into an adequate target version. Ancillary skills
e.g., word processing, Web wizardry, DTPare far less relevant criteria.
This means that graduates fresh from translation schools or universities are
unsuitable candidates. The typical profile of a new recruit is a professional
translator with several years experience either in other organizations or in
3. For example, PECO, a term currently used in official international discourse to refer to the
Eastern and Central European countries (based on a French acronym), was invented by the
OECD Translation Division.

88 The Role of Translation in an International Organization

the private sector. However, OECD records suggest that mature translators
with translation-school training usually make better candidates than mature
translators with only university training in modern languages.
Qualifying External Suppliers. The Translation Division outsources work to
some 100 external suppliers in a typical year. These translators are qualified
on the basis of their personal profiles; there is no qualification examination.
The main purpose of outsourcing translation is to benefit from either specific language-pair skills (e.g., Thai to English), or specific subject-area
expertise (e.g., nuclear-waste management).
Languages Covered. The fundamental mission of the OECD translation service is to translate between French and English, the two official languages,
producing a bilingual version of the knowledge produced by the organization. At the request of the national members, German and Italian are also
provided as target languages. There have also been requests on occasion for
Japanese target translations.
There are no official source languages apart from English and French, but
there may be requests for documents to be translated from any number of
source languages. In a typical year, the Translation Division provides translations into English or French from the standard West European languages
(Spanish, German, Italian), strategic East and Central European languages
(Czech, Hungarian, Polish), strategic Asian languages (Chinese, Japanese,
Korean), and any number of other languages on an ad-hoc basis (e.g., Bulgarian, Macedonian, Turkish).
Since 1985, the relative proportion of French and English target translations
has changed dramatically. Fifteen years ago, 70 percent of all documents
were authored in French for translation into English. By 1999, this relative
apportionment had radically changed: today over 80 percent of all source
documents are in English, and only 20 percent are authored in French.
Target Language

Share of Volume

French

80%

English

17%

German

1%

Italian

1%

Other

1%

These figures clearly show that the L2-authorship phenomenon primarily

affects French translators.

Andrew Joscelyne 89

Total Annual Translation Volume. The Translation Division translated some

80,000 pages (24 million words) of material in 1998 (all languages combined). This represents a 100 percent increase over 1985, or a manageable
10-percent annual increase.
Doc um en t Ty pe s a nd C o nt en t

A translation department in any international organization (or for that matter in any standard in-house translation service, private or public) is
expected to handle almost any kind of document, from half-page memoranda to an occasional 900-page report. The choice as to how a document
translation is handled is a matter of local negotiation with the requesting
service.
The OECD produces a steady stream of knowledge embodied in documents, including:

reports on various issues of interest to its membersfor example, on

social affairs, corporate governance, nuclear-waste disposal, and fishing
rightsranging from 200 to 900 pages and involving a broad range of
technical material;

economic reviews for about 20 of its 29 member countries per year, as

well as twice-a-year forecasts in most economic and business sectors;

official treaties, agreements, and other legally binding documents such

as the Convention on Double Taxation;

conference papers, reports, and summaries of conferences, and other

material associated with international communications on issues of
interest to OECD members.

Documents often include graphs, tabular information, and other formats

depicting quantitative information.
Th e Tran sla t ion P roce s s

In terms of basic workflow, there appears to be nothing specific to the

OECD translation process. A request is received, translated by the person(s)
management assigns, revised, and the target document is returned to the
requesting unit in the appropriate form.
In-House vs. Outsourcing Translation. More significant is the decision procedure whereby a given translation task is assigned either to an in-house translator or is outsourced. The following type of documents point to in-house
translation:

90 The Role of Translation in an International Organization

Confidential. Confidential documents are always translated in-house,

thereby fulfilling one of the key roles of the translation function.

High-Profile. Documents involving the organizations image are also

translated in-house for confidentiality reasons.

Urgent. Speeches from senior offices must be translated rapidly in

keeping with OECD terminology standards.

Committee-Authored. These tend to be hard to read due to the multiple

commentaries and edits that have been added to print documents.
They are often hard to invoice, since they invariably consist of an
already translated version of a previous document scattered with
updates.

Outsourced translation tends to be:

long, nonurgent, lower-priority documents;

meeting reports, where there is a low OECD image quotient and

which are often circulated among insiders;

highly technical documents, where in-house skills do not exist;

more obscure language pairs;

emergency jobs due to in-house sickness, etc.

Revision and Rereading. There are two levels of revision at OECD

instructional revision and rereading.

Instructional Revision. This refers to the work that a senior translator

(i.e., one who has been with the Division for some time), might supply
to a junior (i.e., new) translator who requires aid in learning how to
read through L2-authored or other parameters of OECD documents.

Rereading. A more systematic work-through of the text. This process is

performed either by a member of the Translation Division, or by one of
the requester authors.

In general, the Translation Division is not sufficiently resourced to carry out

a comprehensive revision of all texts. This means that key strategic texts
where the organizations image is at stake are revised, but that other material, of lower priority, is not. This in turn has an impact on whether or not
a document is outsourced for translation. Only translators considered capable of delivering a target document that does not need significant revision
will be chosen.

Andrew Joscelyne 91
Tr ans l a ti o n To o l s and R e sou rce s

Because the Translation Division plays a role in controlling quality of the

knowledge produced in the OECD, terminology and document reference
management is critical to the life cycle of any translation. There are a number of reasons for this:

the high referentiality index: many documents refer to and contain

citations of other documents produced within the same or similar
organization;

multiculturalism of content;

the current lack of a comprehensive translation-memory system;

the relatively wide range of subjects treated in documents.

The Translation Division therefore works closely with the RTU to respond
to these constraints.
Terminology Resources. RTU members are not terminologists in the strict
sense of the word. A certain amount of terminology management (the
extraction, maintenance, and access to terms) is integrated into the overall
workload of the translator.
Translators are expected to use the services of the reference staff when possible to check on documents and validate wording. A translator who spends
an hour on the Internet seeking terms and checking in and out of terminology forums and the like is not rewarded for her/his initiative. On the
contrary, this is considered a waste of resources, since there are staff available
to make more efficient searches for them.
In other words, translation management is organized to ensure that translators work in tandem with reference staff to carry out their special role of mastering document meaning. The driving idea is for all translators to be able
to choose the most effective tools to meet their needs. The RTU team provides translators with documents from a number of text bases, computerized or print. They also manage the aligned bilingual versions of key
resolutions and regulatory documents available online to translators.
Translators themselves have created and maintain a set of about 20 specialist
glossaries on recurring subjects (e.g., nuclear safety, macroeconomics,
e-commerce, agriculture) developed out of their own terminological consultation work based on documents provided them by the RTU staff. These
glossaries are formatted as simple two-column Microsoft Word files and are
available both online and as hard copy.

92 The Role of Translation in an International Organization

The Translation Division has successfully marketed a few of these glossaries

as printed volumes (the first editions of the economics and agriculture glossaries have sold out and revised editions are in the pipeline). The department also makes use of the Microisis bibliographic database system
developed by Unesco as a tool for ensuring that document titles and acronyms are systematically available to translators.
Cooperation Among Internal Organizations. The OECD translation resources
are not limited to in-house material. Unlike most corporate terminology
users, intent on protecting their proprietary terminology, international
organizations have an unwritten policy of coordinating and sharing their
terminology and translation resources. This is because the documents produced by these organizations often cite each other, or deal with similar concepts and usage, and therefore benefit from the extension of terminology
validation to any future translation of such documents.
The directors of the translation departments of international organizations
meet regularly to discuss such collaboration, and there is even a project
under way to make all international-organization terminology available
electronically via an intranet for direct access by translators with access
rights.
In addition, there appears to be a relatively significant flow of staff translators between international organizations. One of the effects of this community of international-organization translators is to call on one another for
terminological and other aid when translating, and thereby benefit from a
hotline to insider knowledge.
This informal dimension of professional translator behavior also exists in
other sectors (for example, all translators who work on medical documents
between certain languages) and has been largely enabled by the emergence
of translator and terminology networks on the Internet. But the critical
dimension in international-organization resource management appears to
be at the level of lengthier citations as much as term strings. If a translator in
one organization has easy access to a document, a translator in a second
organization has easy access as well.
Improving Resource Availability. Despite the available manpower and existing
access to a range of terminology in the semi-informal international organization network, the Translation Division is planning to reengineer its terminology access methods. The current plan is to simplify access to glossaries
by porting them into a Microsoft Access database and designing a single
interface. The user will be able to rank the relevance of the various glossaries when searching for a term and benefit from other functions.

Andrew Joscelyne 93

Interestingly, the department has decided to develop its own terminology

database based on off-the-shelf software packages, rather than purchase one
of the dedicated terminology-management systems now on the market.
Translation-Division managers usually explain this decision by expressing
doubts about sustainable support services from dedicated terminology-system developers. They are also expected to purchase software solutions that
are multi- rather than mono-functional.
Translation Tools and Translator Status. We have seen that translators are valued
within the organization as coordinate (as opposed to subordinate) knowledge workers. Much of their relationship with translation tools is colored by
this status.
Until 1988, translation at the OECD was mostly handled in pre-PC mode
by a two-person team: the translator would dictate and a transcriber would
type and format. The PC is now de rigueur in the department and is systematically used for urgent translation work. But translators may still choose to
translate longer reports by dictation and then have them typed up and formatted separately by transcribers.
Translation-Division management does not feel that the PC has significantly reduced the cost of in-house translation, nor has it radically affected
productivity. Since they see their staff as translation professionals, and not a
mix of secretarial worker, data-input operator, and language professional,
they are concerned with optimizing the conditions under which quality
translation is produced, not minimizing the staff or unnecessarily overloading translators with non-core tasks.
This means, for example, that it is more productive for translators exclusively to translate, and have transcribers produce the numerous tables and
graphs that appear in a report. OECD translators are not paid to master a
graphics package, but they are encouraged to be flexible in their use of
toolsusing a PC or resorting to dictation where appropriate.
Machine Translation and Translation Memory. The OECD Translation Division
does not use machine translation (MT) in any form. There is apparently
fairly strong agreement among the heads of translation divisions at the
major international organizations that MT does not yet offer the required
reliability and quality. This is despite (or because of) systematic use of
Systran as a language-support tool at the EC, and a certain interest by the
UN Translation Division in its possible impact on productivity.
A number of organizations produced internal reports during the late 1980s
when the availability of MT technology first reached public awareness. The

94 The Role of Translation in an International Organization

World Bank actually licensed the Tovna MT system, but apparently decided
it was not a suitable translation tool and never fully deployed it. The general
consensus is that MT will not provide a solution to the foundational translation needs of international organizations for at least the next 10 years.
Translation memory, however, is being actively examined and in certain
cases (e.g., at the EC), deployed in the same organizations. On the agenda
at the OECD is a project to introduce Trados Translators Workbench, a
translation-memory system fed by Multiterm, its terminology-management
component. The goal is to ensure that new translations can automatically
draw on validated terminology, thereby harmonizing terminology across all
documents in the organization.
The Translation Division can also benefit from this validation system by
outsourcing terminology-ready source texts to translators, without having
to spend time preparing extensive terminology sheets to accompany the
documents. If this proves feasible, translation memory will offer the OECD
a tool to address one constraint on its translationthe high referentiality
index.
S U M M A R Y

We can briefly pull together the key points developed in this profile, by listing the ways in which the four constraints on knowledge production in a
context of foundational multilinguality impact decisions in translation management.
Constraint

Impact on Translation

Impact on
Management

L2 authoring

Requires trained readers

Contact with authors

In-house capacity
Careful recruitment
MT unusable

Committee-authored
documents

Requires trained readers

Access to in-house document
cascades

In-house capacity
Reference support
MT unusable

Multiculturalism

Requires trained readers

Controlled terminology

In-house capacity
Reference support

High referentiality index

Access to in-house document

cascades

In-house capacity
Reference support
TM usable

This brief profile leaves many questions unanswered: for example, can editors be employed to standardize the L2-authored texts that are currently filtered by the Translation Division? Or indeed, what happens to these source

Andrew Joscelyne 95

documents when the translated targets have improved on them? In particular, it would be interesting to have more comprehensive statistics on the
nature and source of these L2-authored documents.
This wish will be partly answered in due course; the OECD Translation
Division is to start implementing a new translation workflow-management
system in 1999. This will streamline the translation request and production
process and simultaneously provide a rich stream of data about the who,
what, where, when, how much, and how often of the Divisions translation
productivity. Which means that, like other organizations, public and private, it too will be producing instrumental knowledge about its own
behavior.

L O C A L I Z A T I O N
C A S E S T U DY

Section III: Localizing the Product

Localization at Hewlett-Packards
LaserJet Solutions Group
Karen Combe

S U M M A R Y

Hewlett-Packard (HP) is a large, diverse corporation in which each business

function operates with considerable autonomy. This case study addresses
how a single group, the LaserJet Solutions Group (LSG), has reengineered
its language processes.
The Information Engineering (IE) department in LSG functions as an
internal service bureau, providing English-documentation development and
localization services for LSG products. IE selects suppliers, designs deliverables, manages projects, and delivers final output to its clients. IE must
demonstrate its value to LSG during periodic reevaluations. It is therefore
critical for IEs survival that its supplier program be successful, and be seen
as successful.
In this case study, we examine IEs change from exclusively using an HPinternal localization group in Europe to establishing a viable, US-based,
localization supplier program. We also look at the benefits derived from
extending the HP Waymanagement principles typically applied only
internally within HPto its supplier program.
B A C K G R O U N D

For its initial entry into localization, IE went to an internal HP group in

Europe that provided localization into nine European languages. Contracting to small translation agencies and individual translators throughout
Europe, the group could not be considered a full-service localization
agency that offered complete project-management services. One person
within IE, the localization project manager, supported and coordinated the
effort on the US side.
It was hardly a black-box approach, as the US engineers were quick to
discover. In the normal course of localization, translators need access to
subject-matter experts to clarify text. Translators contacted the US-based IE
localization project manager directly, who answered their questions with
help from the engineers or the technical writers within IE. This coordination required considerable effort from IE.

98 Localization at Hewlett-Packards LaserJet Solutions Group

IE recognized several additional shortcomings in this model:

Because the number of projects and languages was slated to increase,

using a single localization supplier, even an internal one, posed a risk.

The eight-hour time difference restricted real-time communication.

In the early 90s, knowledge of internationalization and localization was

not widespread in the LSG labs or among the content developers. IE
managers felt that a more team-oriented approach to localization
projects between HP and a nearby supplier would facilitate sharing of
knowledge and yield better global products.

The translation coordination effort noted above would soon have

required IE to add staff, as the number of languages and products
needing localization multiplied. The table below shows how staff would
have to be added under the old model as projects increased, as
compared to fewer coordinators needed under the new model of
adopting a local partner.
Old Model

New Model

up to 2 projects

1 coordinator

up to 4 projects

1 coordinator

24 projects

2 coordinators

48 projects

2 coordinators

46 projects

3 coordinators

R E Q U I R E M E N T S

These reasons were sufficiently compelling for the company to adopt a new
model, and IE decided to change to a US-based, full-service localization
supplier. HP managers faced two challenges: 1) choosing the right supplier;
and 2) setting up a program that would get the best performance from the
combined efforts of the supplier and HP teams. The supplier would have to
meet multiple criteria:

financial stability, including a sufficiently large customer base so that HP

would not constitute too much of the suppliers business;

proven multilanguage localization capability, with systems in place to

ramp up new languages quickly;

substantial production capacity in translation, desktop publishing, and

engineering;

project-management systems sufficient to handle very large projects;

strong software-engineering skills;

Karen Combe 99

quality orientation and procedures;

compatible corporate culture.

I M P L E M E N T A T I O N

Choosing a supplier proved simple compared with the actual implementation. After a several-month search, IE chose International Language
Engineering (ILE) of Boulder, Colorado (USA).
The supplier program benefited immensely from application of HPs core
management philosophy. This is particularly significant given the nature of
the translation and localization industry. Although localization is complex
and is in the critical path to launching a new product, language companies
have rarely been considered as real partners in the success of new products.
HPs approach to localization represents, in my view, pioneering work in
the application of real partnering with a language-services provider.
In his book, The HP Way, David Packard explained his approach to management: No policy has contributed more to Hewlett-Packards success
than the policy of management by objective. [...] Management by objective [...] refers to a system in which overall objectives are clearly stated and
agreed upon, and which gives people the flexibility to work toward those
goals in ways they determine best for their own areas of responsibility.1
According to management guru Peter Drucker, in the traditional organizationthe organization of the last 100 yearsthe skeleton, or internal
structure, was a combination of rank and power. In the emerging organization, it has to be mutual understanding and responsibility. Packard notes
that, though Hewlett-Packard is hardly an emerging organization, mutual
understanding and responsibility have been, for many years, key characteristics of the HP style of management.2
IE chose its first and subsequent localization partners very carefully and
then invested time and training in the relationships. Having selected what
IE considered to be the best available partners, those companies were then
empowered to do the best job. HP makes its expectations clear and then
gives the suppliers the freedom and responsibility to reach those objectives;
client and supplier are intended to act as colleagues rather than masters and
vendors.

1. Packard, David. The HP Way. HarperCollins Publisher, Inc., New York, pp. 152153.
2. Packard, p. 154. Druckers quote comes from an interview in the spring 1993 issue of the
Harvard Business Review.

100 Localization at Hewlett-Packards LaserJet Solutions Group

For its part, IE shoulders its own responsibilities in localization: defining the
project, specifying reasonable turnaround for issue clarification, attending to
internal problems, and providing its share to any mutual process improvements. By working in partnership with its suppliers, IE has determined that
it gets more for its moneysuppliers start thinking on HPs behalf, continuously looking for improvements in systems and processes.
Below are two examples of how this supplier program has worked.
Exampl e 1: Establi s hi ng Eng in eeri ng Con fid en ce

The first stumbling block in the IE-ILE relationship occurred over engineering issues. In a project that we will call Kinley (not the real code name),
part of the printer software was supplied by a third party. During localization into nine languages, ILE engineers discovered that the localized
context-sensitive help was broken, although the English version functioned
properly. ILE asked HP what could be causing the problem, and HP in turn
referred ILE to the third-party developer. Meanwhile, because of tight
deadlines, the project was advancing through its normal process. In the end,
ILE and HP engineers together determined the problem and its solution.
However, the delay in resolution resulted in a considerable list of content
issues to be taken care of by translators at the final verification (linguistic
QA) stage, which meant that, because of an immutable deadline, normal
QA tasks got short shrift.
The localized software first delivered to HP contained mistakes such as
duplicate hot keys. Although ILE had helped solve online-help problems,
HP engineers saw only the rudimentary errors and questioned ILEs engineering competence. Although the HP and ILE engineers worked together
quickly to eliminate these low-level mistakes and HP delivered a defect-free
localized product, the confidence problem remained.
Rather than simply casting aspersions, HP was able to apply its empowering
management philosophy. IE managers realized that since ILE had clearly
demonstrated engineering expertise in other ways, there must be some
other forces at work here. Instead of blaming the supplier for sloppy work,
HP worked with ILE to understand the root cause.
The root-cause analysis yielded several important conclusions:

HP gained a greater understanding of the importance of process in

large, multilanguage localization projects. If the designated tasks cannot
be accomplished during the appropriate phase, and therefore extra tasks

Karen Combe 101

accumulate for completion in the final phase, then that final stage will
produce quality problems unless that phase is lengthened.

Failure to resolve serious issues had constituted a threat to quality and

delivery. Project managers at ILE accepted their responsibility to clarify
issues sufficiently to HP so that correct decisions could be made.

HP recognized the importance of getting questions answered and issues

resolved promptly.

The software had basic internationalization errors which the HP lab did
not recognize. Conclusions:

The HP labs needed training in software internationalization.

The labs would need to monitor the internationalization status of

third-party software.

HP engineers had little understanding of the localization process

and the relevance of the QA steps involved. This would be
problematic for future working relationships with ILE or any other
localization supplier.

ILE and IE instituted the following process improvements:

An engineer-exchange program was instituted.

At project kick-off meetings, ILE project engineers would visit HP

and review the software with HP developers. This provided a
platform for feedback to the HP engineer on localization issues,
along with the opportunity to resolve issues early in the project.
ILE engineers would become familiar with the software, learn to
build the product, and discover any idiosyncrasies from a languageengineering perspective.

HP engineers would visit ILE once the project was under way to
observe the localization process. ILE would provide structured
training on localization to the engineers.

Over the course of several projects and through mutual feedback

from HP and ILE engineers, a localization kit, called a Localization
Instruction Worksheet (LIW), was developed and refined. This
LIW contained all the information needed for the localization
supplier to complete the software portion of a project.

102 Localization at Hewlett-Packards LaserJet Solutions Group

This exchange program, started in 1994, still continues today.

Based on their increased understanding of localization process, IE

project managers agreed to respond to project-related questions within
24 hours. Although complete answers were not always available in that
time, HP worked towards the fastest-possible resolution. They
understand that if issues are unresolved within times they agree on with
their supplier contacts, schedules may need to be adjusted or QA steps
added to ensure quality.

ILE provided internationalization training to HP software and firmware

labs. HP developed internationalization standards in their labs, and
provided further training in their own teams. This training, which
began in 1994, continues today for new engineers.

ILE modified its standard process by adding a translator QA for projects

with substantial translation during the verification stage.

IE mustered internal support for these initiatives. The manager of the localization group within IE set up meetings with lab managers to discuss
internationalization requirements, and sponsored training sessions for lab
engineers. Lab managers were invited when ILE presented new internationalization or localization technologies at HP. IE localization managers
included engineers in project post-mortems. The figure below illustrates
the integration of localization project management with internal engineering staff.
LABS

IE

Internationalization and
Localization
Requirements and
Technology
Manager

Software
Engineer

Software
Engineer

Manager

Software
Engineer

Localization Manager

Software
Engineer

Localization
Project Manager

Internationalization &
Localization Training
Project Kick-offs
Visits to Suppliers
Post-Mortems

Localization
Project Manager

Karen Combe 103

Example 2: Outsourcing English D eve lopment

In 1995, IE determined that HP would be best served by outsourcing

development of learning productsthe term it uses for the help and documentation accompanying its hardware and software products. Technical
writers and graphic designers who created learning products had until then
resided within IE. The technical writers also served as project managers on
the cross-functional teams that included software labs, marketing, HPs
localization project manager, and the printing/distribution specialist.
Through closer involvement with localization, IE now understood that the
way English learning products were developed directly affected how easily
they were subsequently localized. For example, if writers index an English
manual by inserting markers and using the index-generation feature in their
desktop-publishing software, translated indexes could be generated with
minimal manual intervention. But if writers used manual techniques on the
English, such as post-editing the generated index, that manual work would
have to be repeated in each localized versionup to 30 languages for HP.
Every error or inefficiency in the original English is magnified many times
over by the time localization is completed.
IE saw the value of combining English development and localization services. HPs suppliers added development resources to accommodate the
new need, and opened local offices for proximity to HPs subject-matter
experts. HP gained the benefits of one-stop-shopping for English and localized languages, as well as the improved service resulting from having local
suppliers.
Once localization and development were combined, other synergies came
to light. English writers could now serve as subject-matter experts for translators and localizers, and could also formulate text more clearly for that
purpose. Writers could provide information for translators before translation began, and answer questions quickly once a project was in flight.
Similarly, translators would provide feedback to writers on the localizability
of the text they were creating, and help raise the level of internationalization in HP products. In addition to outsourcing English development, HP
in effect outsourced localization support.
Within HP, outsourcing English development of learning products evoked a
certain unease, since the help and documentation were viewed as integral to
the products. IE worked to alleviate this response:

For unusual project needs, such as instant response in a time-sensitive

situation, IE implemented special processes to ensure that a suppliers

104 Localization at Hewlett-Packards LaserJet Solutions Group

writer or graphic designer would be available immediately to rectify

any problem.

The suppliers writers and project managers were drawn into HPs
planning process (detailed below).

HP project teams were involved in evaluating bids and awarding

projects, giving them a sense of comfort with a suppliers capabilities.

HP employees worked with suppliers on special projects, such as stylesheet design.

The figure below shows the organizational structure at IE before and after
outsourcing English development.
Information Engineering Department
Old Model

Manager

Project
Planning

Project
Management

Technical
Writing

Graphic
Design

Product
Integration

New Model
Manager

Project
Planning

Project
Management

Project
Integration

Expan di ng the Model

Integrating English development and localization in a single company or

partnership outside HP, compounded with increasingly complex projects,
compelled HP to involve suppliers earlier in project planningin some
cases two to three months before English documentation development
begins, versus the prior norm of two to three weeks. The suppliers writers
and designers give feedback on the technical feasibility of desired options,
serve as experts and consultants to HP project teams, and harness new technologies for information delivery.

Karen Combe 105

Since HP continues to expand into new languages, the suppliers localization experts must validate the English information strategy. For example, a
particular strategy for online information delivery may work well with tools
available for English and even European languages, but not for Asian languages, Arabic, or Hebrew.
Earlier involvement also lets suppliers estimate costs and schedules earlier.
IE project managers may also ask suppliers for cost and schedule scenarios
earlier on, thus broadening their options. HP and its supplier can debate the
pros and cons of various approaches. HP recognized that the stakes were
significant: localization budgets may be four to five times larger than the
cost of English development, and time-to-market for localized products was
all-important.
Re se arch on Early S upp li e r Invol vem en t ( ESI)

In evaluating early supplier involvement (ESI) initiatives in high-tech firms,

researchers at Michigan State University found the following 12 success factors (ranked in order of importance):3
1. The suppliers employees were members of, or participated in, the
buyers project team.
2. There was direct cross-functional, intercompany communication.
3. The two companies shared education and training.
4. The two companies set up common and linked information systems
(e.g., electronic data interchange, CAD/CAM, email).
5. Some of the buyers and sellers personnel were located in the same
place.
6. Technology was shared.
7. Formal trust-development processes were established.
8. Information about customer requirements was exchanged.
9. Information about technology was shared.
10. The two companies shared physical assets (plant and equipment).
11. There were formalized agreements on sharing risks and rewards.

3. Discussed in Jukka Nihtil and Francis Bidault (of the Theseus International Management
Institute of France), Sensitivities of Shared Product Development, Financial Times, Mastering Information Management series, 22 February 1999.

106 Localization at Hewlett-Packards LaserJet Solutions Group

12. There were commonly agreed performance measures.

Establishing trust is a sine qua non of a successful supplier/client relationship:
Trust is one of the most important factors in successful
technological collaborations between companies. Fundamentally it is a belief that another person (or organisation)
will behave in ways that will not be detrimental to us even
if unforeseen circumstances should arise. Yet in this day
and age, when technology is so pervasive, trust may look
like a throwback to a time when ignorance prevailed, a
time when traders had to believe in the honesty and
competence of the other party because they did not have
enough information about them. But now parties to a
transaction have a wealth of information on one anothers
capabilities and reputations (and on alternative partners);
surely information makes trust unnecessary because it can
replace faith with knowledge.
On the contrary, many pieces of researchincluding the
Michigan State University survey mentioned earlier
point to the renewed importance of trust. And this is
quite consistent with what we know of ESI partnerships.
These relationships are quite complex from an organizational standpoint, let alone a technological one.
As companies enter ever more ambitious relationships,
which involve a great number of competencies, to address
markets with innovative, immature technologies, they
expose themselves to a very high level of risk. And trust is
vital in risky situations. Consequently, companies are
treating trustworthy partners as assets of enormous value
because the development of trust between companies is a
lengthy, risky, and thus very expensive, process.4
Although it did not explicitly reference such research in adopting its supplier policies, HP embodied many of these principles in ESI and, as we
attempt to quantify below, was able to reap the benefits in terms of
improved cost-management and faster time-to-market.

4. Ibid.

Karen Combe 107

TO O L S

T O

M A N A G E

T H E

R E L A T I O N S H I P

Relationships as complex as those between IE and its suppliers cannot be

managed without tools and well-defined processes. Two tools proved particularly useful in keeping the companies in alignment.
TQRDCE: A C r it i c al A cronym

HPs most important tool in supplier management is the TQRDCE evaluation, which has the following components: Technology, Quality,
Responsiveness, Delivery, Cost, and Environment.
Although combined HP/supplier teams meet for project post-mortems, the
annual TQRDCE allows IE management to identify trends and summarize
performance. Suppliers are rated on a scale of one (lowest) to five (highest)
in each area, with three defined as meeting expectations. The IE Supplier
Manager receives feedback from each team that has worked with the supplier, and formulates a summary report including areas of excellence and
areas for improvement.
Suppliers generally know how their projects have gone in the course of a
year. But though they may be aware of missed delivery dates or substandard
deliverables, they may not be aware of the overall impact on their HP counterparts. The TQRDCE evaluation is as much a measure of perception as
fact, and perception counts for a lot in relationships. If the delivery of one
project with software problems creates a perception in the HP labs of supplier quality problems, the supplier will understand it must work to change
that perception.
Conversely, if the supplier has discovered areas of innovation that benefit
HP, the TQRDCE offers a medium to call out those successes. If a supplier
can deliver certain languages on an accelerated schedule, for example, IE
may encourage the supplier to institutionalize the process so that HP can
take advantage of early availability of those languages, or extend the innovation to other related areas.
Regular Business Meetings

HPs IE team meets twice a year with each suppliers management team.
Each company updates the other on its strategic direction, accomplishments, and plans for the upcoming period. This information allows the
suppliers to prepare for new requirements, whether they be more micro
issues such as the need to add Lithuanian translation capability, or more
macro issues such as the need to cut project cycle time. HP is made aware of
its suppliers emerging areas of expertise or expanded service offerings. The

108 Localization at Hewlett-Packards LaserJet Solutions Group

contact between management teams helps to keep the partners in strategic

alignment and enhance the relationship.
In addition to the management-level meetings, HP or the suppliers may
initiate meetings to address more tactical issues such as refining project estimating or streamlining the review of translations by HPs in-country
representatives.
B E N C H M A R K I N G

In the critical early years following the change to US-based suppliers, IE

measured program success in three key areas: quality, cost, and on-time
delivery.

During that period IE did not submit localized products for in-country
review on a per-project basis. Instead, the product champions in each
country were polled annually. The positive feedback from these experts
assured IE that suppliers translation quality was good.

Cost comparisons over several years and across multiple projects

indicated no increases in unit costs for translation/localization, while
volume of work increased and cycle times were shortened.

Careful tracking of projects showed improved on-time delivery with

the new program.

IE continues to benchmark its program internally based on the hard criteria

of cost, on-time delivery, and quality. Suppliers generally have lowered
prices with heightened efficiency, and time-to-market pressures have also
forced shorter turnaround times. In-country reviews are now performed
more frequently, particularly for the recently added languages, so that HP
gets regular project-level feedback on terminology and translation quality.
IE has also added some softer, less tangible measures:

Supplier Evolution. Client requirements are continually changing;

suppliers are expected to design and implement improvements across
the whole range of 30 languages covered.

Productivity Trend Data. With time, project budgets should decrease

comparatively, as more text is reused and processes are improved. IE is
finding its own headcount is staying relatively flat although it is
handling more projects.

Perceived Value. IE gets regular feedback from its internal customers on

the value of the services they provide to LSG.

Karen Combe 109

C O N C L U S I O N

By becoming students of the localization industry, IE managers have been

able to assure that their translation program is competitive. They attend
industry conferences, such as STC, LISA, and Seybold, and monitor industry-watchers such as The Gartner Group.
HPs IE group has implemented an English development and localization
program that responds effectively to the needs of internal customers and
ensures IEs value to the LSG organization. By applying HPs internal management philosophy to the supplier program, IE has developed partners that
work on their own and on HPs behalf to continuously improve process,
increase value, cut costs, and decrease time-to-market. This case, from both
HPs and its suppliers perspective, has all the earmarks of a success story.

L O C A L I Z A T I O N
C A S E S T U DY

Shortening the Translation Cycle

at Eastman Kodak
Suzanne Topping
O V E R V I E W

The worldwide market for digital cameras is among the most competitive.
The hottest sales period for new models typically comes in the first weeks
following product introduction. And although Kodak tried to strategically
schedule announcements around the world, once a product appeared in the
US, news spread quickly. Kodak needed to get the cameras into other
regions sooner to take advantage of the initial buying flurry.
Translation of product literature and software is, by definition, in the critical
path for shipment of these products: no translation, no product. Marketing
staff, not familiar with translation and localization complexities, saw the
typical timetable for language work and asked, predictably, Why does
localization take so long?
Typical Timeframe for Release of Localized Products
English/Japanese

English/Japanese/
French/German

English/Japanese/French/
German/Simplified Chinese/
Traditional Chinese

In boxes day 1

In boxes ~4 weeks after CD #1

In boxes ~4 weeks after CD #2

The marketing department was intent on cutting localization time, and targeted simultaneous availability of the product in all languages (not the
current scenario of the foreign-language versions lagging the English).
While the lack of understanding on the part of the marketing managers was
frustrating, we in the localization group understood that their need to get
localized products sooner was based on hard market drivers.
Process improvement and cycle-time reduction had been an ongoing initiative of the localization group. Marketings request for simultaneous launch
without English delay simply escalated the need for change. The original
request for simultaneous shipment came in September, 1998, and the first
successful simultaneous launch occurred in April, 1999. While many of the
product-development changes and evaluation of QA process alternatives

112 Shortening the Translation Cycle at Eastman Kodak

were already underway, the bulk of the research and the actual changes all
took place within a seven-month period.
WO R K L O A D

S T A T I S T I C S

The Digital and Applied Imaging localization team handled a variety of

project types and sizes throughout the course of the year.
Sm al l Proj ects

Type. Project types ranged from a few strings to small packages or flyers.
Size. Up to a few hundred English words. Vendor costs were generally less
than US$1,000. Localization-management time was between one and
five hours.
Volume. About four small projects per month.
Language sets. Language requirements could be as simple as a single language
(like Japanese), a few languages, or a wide range of languages (e.g., French,
Italian, German, Spanish, Japanese, Simplified Chinese, Traditional Chinese, Russian, and Netherlands Dutch).
Turnaround times. Generally quite short. Very little advance notice about the
project was given to the team (under a week), and clients usually needed
them within a few days. Project timelines were determined by clients.
Elapsed project time was consistently under two weeks, and usually less than
one week.
Resources involved. Resources required for small projects were limited. Typically, the jobs were for a single person (e.g., a package designer). One
localization team member managed the project, using a multilingual vendor
to get the translation done. Sometimes a second vendor was used for linguistic quality assurance (QA).
Mi d- S ize Proj ect s

Type. These ranged from individual users guides to product maintenance

releases. Maintenance releases could involve:

changes to software only (in all languages);

addition of a single language to the existing product language set

(impacting all components);

changes to a product name (impacting all languages and all

components).

Suzanne Topping 113

Size. From several hundred to 50,000 English words. Vendor costs were
generally between US$1,000 and US$20,000. Localization-management
time was between five and 20 hours, over the course of four to six weeks.
Volume. Ca. six per year.
Language sets. As few as two languages or as many as five (typically French,
German, Japanese, Simplified Chinese, and Traditional Chinese).
Turnaround times. More advance notice was typically given for mid-size
projects. Planning might begin up to a month before localization activity
started. Turnaround times were usually short, however. Software changes
were frequently completed within two weeks. Users-guide revisions could
take up to four weeks. Turnaround times were negotiated between the
client and vendor. Elapsed time for entire projects was between four and
six weeks.
Resources involved. This depended on the project. For software changes,
players were the localization project manager, a software lead, up to five
software engineers, an installer developer, a QA lead, a few software testers,
and a help-system developer. Multilingual vendors did the translation, and
occasionally a second vendor provided linguistic QA. Regional QA
reviewers, print production staff, and logistics personnel were also involved.
A typical project of this type involved approximately 13 Kodak productteam members, along with vendor resources.
L ar ge P ro j e c ts

Type. Typically localization efforts for entirely new products. Required

localization of all components including: camera or scanner hardware
LCDs, multiple pieces of software, help files, readme files, documentation,
and packaging.
Size. From 50,000 to 100,000 English words. Vendor costs were generally
between US$20,000 and US$200,000. Localization-management time was
close to 200 hours, over the course of about four months. The percentage
of time spent in various project stages is listed below.

114 Shortening the Translation Cycle at Eastman Kodak

Month

Percentage of Localization
Management Time

5%

10%

Project Execution

80%

Wrap-Up

30%

Stage
Planning

Volume. Ca. five large projects per year.

Language sets. Same as for mid-size jobs.
Turnaround times. Extensive advance planning took place for large projects.
Planning would begin three months before localization started. Turnaround
times remained aggressive. Software localization took place in several iterations over the course of about three weeks. Users-guide localization was
scheduled to take six weeks, but usually took closer to eight.
Resources involved. Many internal resources were required for large projects.
In addition to the typical players in mid-size projects, large projects required
a package designer, a user-document author, and a product-marketing manager. One localization staff member managed each project, with support
from the teams manager. Multilingual vendors were used to do the work,
and occasionally a second vendor provided linguistic QA. Regional
reviewers, print-production staff, and logistics personnel were also involved.
A project of this type might involve some 27 Kodak product-team members, along with vendor resources.
Deta i led Ca mera -Product Descri pti on

Digital cameras were the most complex large projects needing localization.
Since they were consumer-oriented, the products included substantial software elements, and packaging and presentation were glossy and rich in text.
Typical components included:

Camera user-interface (fewer than 200 words)

Packaging (1,261 words)

Users guide (22,415 words)

Quick-start card (856 words)

Miscellaneous printed pieces (3,000 words)

Software (3,370 words)

Suzanne Topping 115

Help systems (30,705 words)

Readme files (16,801 words)

Installers (2,920 words)

CD sleeves (96 words)

CD silkscreens (10 words)

The software itself consisted of multiple Windows applications and two or

three Macintosh applications. Additional prelocalized applications were also
included on the CD, and installed by the same installer. All software for
both platforms was presented on a single multilingual, multiplatform CD.
Simplified Version of a Typical Camera Offering

Camera LCD

Marketing
Envelope
CD
Sleeve

User's Guide
Software CD

Packaging
(all exterior faces containing text)
Miscellaneous Nonlocalizable Items

Windows
Installer
Application 1
Help System 1
Application 2
Help System 2
Application 3
Help System 3
Readme
Readme
Readme

Macintosh
Installer
Application 1
Help System 1
Application 2
Help System 2

Geographically Dive rse Te ams

As the product itself was complex, so was the development-team structure

to manage it. Kodaks digital-camera teams were geographically diverse,
with team members in Rochester (New York), Boston, Yokohama (Japan),
and Harrow (UK). For some teams, even project management was divided,
with co-project managers in Rochester and Yokohama (where the cameras
were manufactured).
Localization was primarily managed from Rochester, but most Japanese
items were created by Kodak Japan, and some of the software was handled
in Harrow.

116 Shortening the Translation Cycle at Eastman Kodak

Breakdown by Site of Localization Product Development

Rochester
E

G C

Boston
E

Japan
C

Project Management
Localization
Management
Software Development
Help System
Development

Installer Development
Software QA
CD Builds
User Document
Development
User Document
Printing
Packaging Design
Packaging Printing
CD Sleeve Printing
Product Pack-out

U.K.

As the figure above shows, packaging design and users-guide writing took
place in Rochester; help systems were developed in Boston; software was
developed by a variety of engineers in each location.
Both the Rochester and Boston teams conducted software and CD testing.
While the Rochester site also managed the production of CDs for all countries, the CDs themselves were sent to Yokohama for packaging with
Japanese products. Rochester also managed the printing of English, French,
and German documents, while Yokohama printed the Japanese and Chinese materials.
Having a distributed team meant that information and files came from
diverse locations to be localized centrally, and were returned again once
localization was complete. Tracking each item throughout localization and
production cycles was crucial.
Team Communication

The physical distance between team members complicated communications, which were vital on projects with such tight deadlines. Weekly
product-team, software, localization, and logistics meetings were held in
Rochester, with team members calling in from remote locations. Videoconferencing was used for periodic reviews. Because of time differences,
meetings were frequently held either evenings or early mornings US
Eastern time. Email remained the primary method of communication; it
was used to broadcast messages to team members, verify status of action
items, address or raise issues, distribute status reports, and create a paper
trail for various events.

Suzanne Topping 117

D I A G N O S I N G

T H E

P R O B L E M

In reality, translation and localization itself actually took only a very small
fraction of the total time. The figure below illustrates the activities and
approximate duration for each task associated with producing localized
CDs. Only a few days were spent updating localized versions, and most of
the time was spent on testing and production.
Breakdown by Task of Localization Process
In boxes day 1

In boxes ~4 weeks after CD #1

In boxes ~4 weeks after CD #2

English/Japanese

English/Japanese/
French/German

English/Japanese/French/
German/Simplified Chinese/
Traditional Chinese

Localization
begun on Beta
versions of
English for all
languages.
Vendor testing
performed on
these versions.

Gold English to Vendor

Vendor Debugs Build
Problems (2 days)
Vendor Updates Beta F/G
(3 days)
"Final" F/G from Vendor
Build CD (1 day)
Acceptance Test CD (4 days)
Report Bugs
Receive Fixed Items (2 days)
Build CD (1 day)
Acceptance Test (3 days)
Declare Gold Master
Duplicate CD (5 days)
Pack Boxes (2 days)

Vendor updates beta Cs/Ct

(3 days)
"Final" Cs/Ct from Vendor
Build CD (1 day)
Acceptance Test CD (3 days)
Report Bugs
Receive Fixed Items (2 days)
Build CD (1 day)
Acceptance Test (2 days)
Declare Gold Master
Duplicate CD (5 days)
Pack Boxes (2 days)

The people in the trenchesproject managers, software developers, QA

testers, and documentation specialistsunderstood that the entire productdevelopment process itself needed to be restructured to achieve simultaneous
worldwide launch without English delay.
Various digital cameras had been localized several times already, and the
team had made impressive strides in cutting cycle time for printed documentation. Packaging, users guides, and other printed materials were
already on an accelerated track (frequently underway even before the actual
software-user interfaces or camera LCDs were finalized)and the translation time saved proved invaluable in accommodating long print lead times.
With the process for printed material in reasonably good shape, the biggest
challenges were related to translating and localizing the CD-based software.
In the end, the entire software-development process was examined and
refined. It was broken down into three stages:

Software development

Translation and localization

118 Shortening the Translation Cycle at Eastman Kodak

Multilingual software QA and testing

Since translation and localization time was already quite short, the greatest
savings would eventually come in development and testing.
Problem Area 1 Softwa re D eve lopm e nt
Software-Handoff Problems

Since the digital cameras included so much software and other soft files,
the absence of a strong procedural link between developers and localizers
caused numerous difficulties in creating multilingual CDs. Further accentuating the problem, software came from a variety of locales and sources. And
while some developers structured their applications well and handed off
complete packages for localization, many did not.
For each project, the localization team gave blank specification forms to the
software developers, hoping that these checklists would remind developers
to provide all required files, graphics, environment information, and build
instructions along with the resource files. The team provided the completed
checklists to the localization vendors, along with the files. In most cases,
however, the vendors still did not receive everything they needed; build
instructions were incomplete, header files were missing, or tool versions
were listed incorrectly.
Tracking down answers to vendor questions became a seemingly infinite
loop. Once one question was answered, it frequently opened the door to
another problem. For example, the localization vendor might try to run a
build, only to find that a header file was missing. Once the header file was
sent to them, they would run the build again. The compiler would go one
step further, but stop to report that another header file was needed.
Working through these issues frequently delayed the start of localization by
several days. When dealing with timelines as tight as those for the digital
cameras, these delays are enough to derail a schedule.
No Identification of Strings

Kodak commonly received software back from vendors which contained

strings (such as registry entries) that were translated when they shouldnt
have been, and other strings left in English when they should have been
translated. This occurred because no instructions were given about which
strings to translate. The vendor had to sift through resource files, trying to
figure out which strings in the code were translatable.

Suzanne Topping 119

Complex Build Processes

The process of building software tended to be complex, requiring shared

files and structures provided by other software developers. For example, one
application needed a COM .dll to run, and that .dll was being built with
the rest of the application, and not merely provided for the applications use.
Handing off such a complicated build process to a vendor was inefficient.
Undocumented Build Processes

For several years, software developers had been working under tight deadlines, leveraging applications over and over again. This meant that they did
not tend to document how they built their applications. Unfortunately,
localization vendors were flying blind, and the developers were not very
thorough when it came to passing along instructions at localization time.
Readme File Formatting

A surprising number of problems came from an unexpected source: readme

files. Part of the problem was that there were so many of them, covering a
wide range of issues. For one camera project, there were a total of six
English readme files. When five new languages were added, the result was
36 opportunities for something to go wrong. The most common problem
found when receiving translated readme files from vendors was corrupt
format. This was caused either by using a tool other than the one the
English author had used, or by using tabs and soft returns rather than spaces
and hard returns. The group also found that opening a file in various text
editors yielded different formatting results, even when the same formatting
approach had been applied.
Sol uti ons

The localization team held brainstorming sessions with software developers.

One outcome of these sessions was a key procedural document, Development Guidelines for Easy Software Localization. The information in these
guidelines seemed merely common sense to the localizers, but writing it
down seemed to formalize and legitimize it. And since the developers
themselves suggested the solutions, there was immediate buy-in from them.
Some of the most significant changes were to:

Isolate localizable strings. For each application, all translatable text would
be pulled from the various resource files and placed in a single, textstring-only .rc file.
Handling the text in this way made instructions completely clear for
the vendor. Everything needing translation was in one file, and no

120 Shortening the Translation Cycle at Eastman Kodak

other files should be touched. It also helped prevent accidental changes

to functionality, since the translated file did not contain code.

Make all font information modifiable. For some applications, font

information was defined in code rather than in editable resource files.
This prevented the localization vendor from making adjustments to
increase legibility or to fix font-related bugs. The developers agreed to
ensure that all font definitions were editable during localization.

Simplify build processes. The developers agreed to review their build

processes and pare them down to a more elegant form.

Document build processes. The development community agreed to start

documenting how builds were run. Explicit instructions would be
written so that anyone could sit down at a PC and build the application.
Not only did this help localization, but it also made it easier to
transition work between developers.

Pre-build before handoff to the vendor. As a final precaution, someone other

than the developer would build the software before it was sent to the
vendor. This meant that missing files or other problems were identified
before localization actually began.

Deliver readme files as PDF files. English readme-file authors would use
word-processing or desktop-publishing applications rather than a text
editor, and use all of the formatting tools available (tables, bolding,
etc.). Once files were completed, they would save them as PDF files,
and translated readme files would be delivered in PDF format as well.
This precluded format changes that arose depending on the text editor
in use.

Most of these principles were included as requirements in software-specification documents, to help ensure that the guidelines were followed.
Problem Area 2 QA/Testing
Back-End Problems

Once the front end of the process had been addressed, it was time to look at
the back end.
The QA process faced far more significant challenges than development. As
in most software communities, QA always pays the price for schedule slips
and feature creep (changes in features while the project is being finalized). No matter how significant the changes to plan, QA had to struggle to
somehow make up for the lost time. In addition, QA groups were often

Suzanne Topping 121

strapped for resources, and frequently had to rely on inexperienced contract

labor. (Total number of testers was between eight and 10, depending on the
project.) Engineering releases for existing products tended to keep the QA
leads busy between peak camera-project timelines. Because of these factors,
test plans and cases were frequently not kept up-to-date, and by the end of a
project, they often did not match the features of the product. The QA leads
were aware of this problem and worked around it with the testers, since
they didnt have time to update the documents.
Test labs were also a problem. While language versions of the various operating systems were available, there were not sufficient systems to run several
languages under the same configuration. This created troubleshooting slowdowns, because developers reading bug reports frequently did not know if
the bug occurred in all languages, each operating system, or in more than
one camera model.
The development of the DC220/260 digital camera in early 1998 illustrates
the challenges that these products presented. The cameras were launched at
the same time, and used the same software CD and users guide. They had
slightly different feature sets, however, and the full suite of test cases had to
be run on both camera models. The figure below illustrates the test scenario
for one of the cameras. Before shipping the multilingual CD, this testing
had to be run on both cameras.
Test Scenario for DC220/260 Camera
DC220/260

Windows95

Windows98

Macintosh

122 Shortening the Translation Cycle at Eastman Kodak

For Windows systems, each set of file folders in the figure represents an
installer, several applications, matching help files, and four readme files. For
the Macintosh, the folders represent a slightly smaller set of files. (Cameras
after the DC220/260 would also support Windows NT, and would include
more applications for the Macintosh.)
Given the large number of camera/application/platform/language combinations, testing all languages simultaneously would require a huge lab and
many test engineers. Since testing occurred only at project end, or roughly
every six months, it was hard for Kodak to justify investing in these
resources.
Sol uti ons
Changing the ModelOutsourced vs. Internal Testing

As described earlier, testing was handled in a distributed fashion. Functionality testing was conducted by groups in Boston and Rochester, and
interoperability testing (of English software only) was conducted by a separate group in Rochester. The first task, therefore, was to evaluate whether
or not these internal organizations could manage simultaneous testing of all
languages.
It was fairly quickly determined that this approach was not feasible due to
the following issues:

Camera Availability. If testing were distributed between several Kodak

sites, each would need a minimum number of cameras. But there were
insufficient preproduction cameras for each site; centralized testing
would require fewer cameras.

Lack of Linguistic Resources. Test engineers within Kodak were primarily

in the US. Some might have a passing familiarity with another
language, but this was inadequate.

Difficulty in Obtaining Localized Third-Party Applications. One of the

advantages of using a localization company for testing was the
availability of localized versions of operating systems and applications
(e.g., Simplified Chinese versions of the graphic software Adobe
Photoshop). Further, copies of localized software are expensive and
often difficult to obtain in the US.

Testing Inefficiencies. Distributing testing by language within Kodak

would prevent the simultaneous comparison of bugs. However, if
testing were split up so that each site tested a few applications in all
languages, there would be no efficiencies in minimizing the number of

Suzanne Topping 123

languages/systems that would need to be maintained in each site.

Neither option was attractive.
Once the team determined that outsourcing was the only viable alternative,
the next step was to evaluate vendor options. Throughout the previous
year, many localization companies had begun offering testing as a separate
service in light of perceived market need. To ensure a thorough evaluation,
the localization team involved QA and software development in qualifying
vendors.
Flow Chart of Vendor Selection Process
Start of
Evaluation

Questionnaire #1

Vendor
Disqualified

Pass

Phone/Email
Interviews

Vendor
Disqualified

Pass

On-Site Visit
(Questionnaire #2)

Vendor
Disqualified

Pass
Vendor
Selected

To gather information quickly and objectively, a series of questionnaires was

developed. The first requested information on quality and process metrics,
processes, and sample documents. A second was designed for use during
on-site visits to the vendor. The goal of this questionnaire was to investigate
issues of particular concern, and to ask for demonstrations of technologies
and evidence of required resources.
After the list of potential vendors was narrowed, on-site visits were made to
five promising companies. The primary goal was to find a company offering
testing services in the United States, but companies in Ireland were also
investigated.
Managing the Change

The team realized that for outsourcing to work, the entire testing paradigm
would have to change. Some of the most significant issues identified are
listed below.

Software specifications had to be thorough, and finalized early.

124 Shortening the Translation Cycle at Eastman Kodak

QA plans and dozens of test cases had to be updated throughout the life
of the project, to ensure that changing functionality was accurately
represented.

Kodak, the localization vendor, and the testing vendor had to agree on
terminology before testing could begin. Therefore, glossaries had to be
developed, reviewed, and approved early.

Clear delineation of responsibility had to be established between what

the localization and testing vendors would do. The localization vendor
would only perform basic user-interface testing, and the test vendor
would only perform functionality testing.

Localized-software testing would begin on beta localized versions,

rather than on the versions built from the final English software.

Localized-software testing would run in parallel with the English

testing, rather than wait until English testing was complete.

Kodak QA staff had to evolve from doers to managers of vendor

activities.

Kodaks localization project management and its two vendors had to

develop a communication process for bug reporting so that information
would flow smoothly.

Given the short timeline and the distributed development group, accurate
identification of who owned each bug was crucial. Determining whether a
bug was caused by localization or by development had to happen very
quickly. Communication was critical.
Communication Flow During the Localization Process

Kodak

Localization Vendor

QA

Localization
Team

Test Vendor

Development

Suzanne Topping 125

R E S U L T S

The efforts of the team eventually paid off. The DC240 launched in spring,
1999, with all foreign-language versions shipping concurrently with the
English.
One problematic area remains, related to version control during localizedsoftware testing. The team found some confusion when managing software
versions being tested by the testing vendor. Occasionally, bugs were
reported that had already been corrected. For future projects, version control and communication related to version levels need to be addressed.
S U M M A R Y

Short time-to-market cycles for innovative products are critical for success
in todays competitive business environment. For Kodaks digital cameras,
shortening localization time has meant revamping product development.
As long as the time spent on localization is greater than zero, there will
always be pressure to do it faster, better, and more cost-effectively. We who
have been instrumental in translation and localization at Kodak are prepared
to meet these daunting yet exciting challenges in the new millennium.

L O C A L I Z A T I O N
C A S E S T U DY

Making a Global Product

at MapInfo Corporation
Ricky P. Thibodeau
MapInfo Corporation

In recent years, software sales outside the US have outpaced domestic

growth. US software companies often report international revenues exceeding 50 percent of total sales. One contributor to this growth is software
localization.
Creating a version of a program that is appropriate for a group of users in a
particular part of the world entails not only translating text into the target
groups language, but also recognizing relevant aspects of the local culture.
Since localize means to adapt for a specific locale, it is more accurate than
translate. Localization entails the linguistic and cultural adaptation of the
software and user-interface (UI), documentation (books, online help, tutorial), and packaging to accommodate requirements in foreign markets.
In this case study, we examine how MapInfo, a fast-growing US company,
developed a successful localization strategy for worldwide distribution.
L O C A L I Z A T I O N

O V E R V I E W

A major reason for localizing software is economic. Initially, software localization may increase a companys sales by at least 25 percent. A product that
is barely making a profit domestically can be a highly profitable venture
overseas. All else being equal, a software product that is not localized is less
likely to survive over the long term.
A localized product will help spread R&D dollars over a wider base, as a
localized version can extend a products life cycle. If the domestic market is
declining, an international market may be emerging or still growing, and
sales abroad will help finance the next generation of products. In addition,
most products can be more profitable overseas, since these markets often
support a higher price level.
A localization strategy outlines the goals for product localization and helps
determine which products to localize and which markets to enter. The
strategy should fit into the companys overall sales and marketing plan to
ensure that products planned for localization will enable the sales organiza-

128 Making a Global Product at MapInfo Corporation

tion to penetrate the desired markets. Decisions can then be made about
what languages to choose and when to release each language edition. Based
on this strategy, some products may not be localized at all. At MapInfo, we
use business models and do careful analysis for each language to determine
if localization is warranted.
Val ue o f C o mmu nic a t ion

Successful localization requires an entire organization to embrace the idea

that its products are being designed for international markets, and to set up
the appropriate procedures and channels of communication.
From the very beginning of the product-development cycle, communication channels need to be established with key people in each department.
Product managers must bear in mind that each proposed product may also
be marketed internationally. Developers should both understand and
observe internationalization requirements while writing code. Documentation material must also be developed accordingly. Quality Assurance (QA)
testing needs to include an international plan. Finally, shipping must be well
coordinated with the overseas operations. When all key team members of
product development proceed with international considerations in mind,
localization can be successful.
Mod el s

There are a variety of models that a company can use to implement a localization strategy, and it is imperative that a company choose a model that fits
within its long-term goals. Below we outline the most common models
used in our industry.
Distribution Network

The distribution-network model is often used for new and emerging markets, but serves equally well as a main global strategy. In this scenario, a
company already has a worldwide distribution network in place. These distributors are then enlisted to participate in the actual localization of the
product. The main advantages are that distributors already have a strong
product knowledge base and an understanding of corporate plans, as well as
low overall localization costs. Conversely, the distributor may not have the
resources in place to meet schedules for deliverables, thereby prolonging
ship dates.

Ricky P. Thibodeau 129

Single Language Vendor (SLV)

With the single-language vendor (SLV) model, localization is usually done

in-country with native translators who can naturally observe the cultural
requirements of the target language. This model can be used when testing
markets and is more targeted than the distributor model. SLVs are more
likely to meet localization schedules for deliverables, thereby allowing international product shipments to occur shortly after English product shipments. In the long run, the SLV approach may be more economical than
the multiple-language vendor (MLV) approach discussed below.
Multiple Language Vendor

The multiple-language vendor model may be a good alternative, particularly if the goal is simultaneous release in multiple languages. In this
approach, each project is supervised by a team leader who assembles a team
of project managers, translators, editors, and proofreaders to coordinate and
ensure a quality translation for each project. Centralization of product management can lower costs and raise consistency and leveraging. This model
may not be cost-effective for smaller companies.
Mixed In-House/Outsourcing

The mixed in-house/outsourcing model is probably the most popular and

flexible, since the publisher decides the strategy for each project. However,
localization activities scheduled in-house may not be a viable model for
smaller companies since they would need to secure in-house resources to do
the localization during product development.
L O C A L I Z A T I O N

A T

M A P I N F O

MapInfo provides Internet-enabled products and solutions for businessintelligence applications. By detecting and charting patterns in corporate
databases, MapInfo products provide greater overall corporate efficiency.
Over 300,000 users in a wide range of industries use MapInfo solutions to
streamline business activity. Its products are deployed on the desktop, on
enterprise servers, on the Internet, and as embeddable mapping objects in
decision-support and enterprise-wide applications.
Headquartered in Troy, New York, MapInfo has over 450 employees
worldwide (295 in North America). MapInfo products are available in 21
languages and are distributed in 58 countries by more than 1,000 partners.
The business has grown anywhere from 10 percent in the more mature
markets of Australia and New Zealand to 31 percent in Southeast Asia and
Europe. Direct sales versus indirect sales varies geographically, except for

130 Making a Global Product at MapInfo Corporation

Southeast Asia, where all business is conducted through channel partners

and local distributors.
MapInfos main European office is located in the United Kingdom, with
Asia/Pacific support offices in Australia and Japan. We also have offices in
Germany and Sweden, as well as a strong distribution network worldwide.
The operations manager in the UK office is a key figure in overseeing the
success of our European localization activities. Similarly, our corporate
Asia/Pacific team and our Australian office are critical for our success in
Asia/Pacific.
Upon deciding to implement localization at MapInfo in 1995, we concentrated our efforts on MapInfo Professional, our main desktop product,
localizing it into 13 languages worldwide. This product is now available in
21 languages, including English, generating over 50 percent of our revenue
in foreign markets. We have also started to localize many of our other products, including MapX (developer product), MapXtreme (Internet/intranet
product), and SpatialWare (enterprise/client-server product).
Local i zation R ole s
Role

Responsibility

Localization Director

Manages the entire localization process (includes localization

agreements, finalizing localization schedules for each product
release, coordinating activities with international offices, and
overseeing staff to meet corporate goals and objectives)

Software Engineer
(Speaks Fluent
Japanese)

Internationalization-related activities (includes automating

localization-kit builds, verifying localization kits, developing
international Localization Automated Testing System [LATS],
managing localization schedule, project managing for Japan,
fixing localization bugs)

Localization Engineer
(Understands French,
Spanish)

Builds of each language, Notes administration, installer issues,

assisting with localization-kit builds, coordinating
documentation, testing language builds, debugging, beta program

Localization Engineer
(Speaks Fluent
Mandarin)

QA lead, assisting with ATS, project managing for China/Korea,

providing technical support for Asia/Pacific, testing of each
language edition, fixing localization bugs

International Team

Europe and Asia/Pacific support (includes production, shipping,

selecting vendors/localizers, localization agreements, budgets)

Engineers

Internationalization requirements, fixing bugs

QA Team

Running automated localization-kit builds, international testing

plans, investigating bugs related to English products

Localizers

Localization, local testing, overseeing local beta program

Ricky P. Thibodeau 131

MapIn fos Model

Localizing MapInfo Professional primarily entails the localization of books,

online help, and softwarebut there are dozens of additional elements to
manage, from sample data to media-label art. The NY-based R&D localization department distributes the English resource files (menu items, error
strings, dialogs, bitmaps, etc.), books, and online help electronically to our
international localizers, who are largely our international distributors. The
localizer performs the translations and returns the localized files back to us
via the Internet translated, compiled, and tested. These files are then substituted for the equivalent English files to produce the local-language editions
of MapInfo Professional. We then return complete, installable image files to
each localizer for final testing. We work with the distributor, our production department, and our international offices to coordinate duplication,
assembly, and fulfillment of the full products. A sample localization-kit
directory is as follows.

Notice that we distinguish the user-interface (UI) resources from other

resources. This helps localizers identify which components are needed to
compile the resources from other support files such as documentation and
installer resources. We separate our localization of 20 languages into the following two groups:

Full localization (14 languages). This includes the UI, books, online help,
online tutorial, and third-party applications. Full production of the
language-specific product is done. Full localization is necessary in the
large European and Asia/Pacific markets.

Partial localization (6 languages). The UI is localized and occasionally

online help is as well. Distributors ship our standard English version and

132 Making a Global Product at MapInfo Corporation

include a diskette containing the localized resource DLLs and support

filesa cost-effective plan when a company wants to enter a smaller
market without incurring high localization costs associated with
documentation. For MapInfo, these language markets include Arabic,
Hebrew, Czech, Slovak, Dutch, and Turkish.
Prior Model

Early on, our strategy was to use MapInfo distributors as localizers. Our
country managers from Europe and Asia first identified which distributors
had adequate resources. MapInfo would then initiate a localization agreement, compensating them with sales credits. Very rarely did a distributor
that we approached turn down the opportunity to participate, since they
wanted to ensure product quality. Also, this model provided them access to
corporate staff such as engineering, technical support, and quality assurance,
via the localization department.
All 13 languages that we shipped in 1995 were localized by our distributors,
with the exception of Japanese. Using the distributor model, MapInfo saved
hundreds of thousands of dollars. The Japanese language was done by a very
large MLV, costing six times more. This first foray into the MLV process has
deterred us from using MLVs, although we do not rule out future work
with them.
Current Model

Such cost-effective localization requires strong project management and

communication with distributors and localization vendors. We use Lotus
Notes, further explained below, as our main tool of communication with all
localizers.
Currently, we still rely heavily on our distributors. We use 14 MapInfo distributors and six localization vendors, all of whom are SLVs doing the localization in-country. This hybrid model (distributor/SLV) works well. SLV
companies tend to be better than distributors at meeting deadlines.
As our core product now includes multiple features from third parties,
increasing the localization work, we have implemented a model with three
of the languages whereby some distributors only localize the UI, given their
strong product knowledge. Documentation (books, help, tutorial) is outsourced to an SLV, freeing up the distributor for marketing and sales,
thereby increasing revenue while still meeting delivery dates.

Ricky P. Thibodeau 133

Product Deve lopment with Emphasis on Internationalization

The localization team is very much involved with product development

from the beginning of a project, and globalization is at its core. Although
developers may realize that the product being built will be localized, it is
still necessary to continually reemphasize internationalization standards.
Some of the major standards we emphasize at the software-architecture
stage include:

writing software with localization in mind: dialogs, menus, text, and

other localizable components should be stored in separate resource files
(resource DLLs), not hard-coded into the source files for the
executable;

sort order and searching differences (e.g., Danish and Swedish

alphabetize differently than English);

double-byte enabling the code;

testing source versions internationalization features;

requiring that third-party components be internationalized.

We usually assign a project manager to each product being built to ensure

that all localization requirements are met. At MapInfo, product development and internationalization are so tightly integrated that they can be considered a single process.
Ma nagi ng Thi rd Pa r t i es

With the high costs of development, it is not surprising that third-party

products are making their way into a companys core product line. When
bundling an application with a core product, the product manager must
address internationalization/localization needs. Clear communication
between the product managers, development, and the localization department is crucial.
Internationalization and Localization Requirements

During a release of our core product, MapInfo decided to include a new

feature developed by a prominent third-party software company. Corporate
Engineering and Product Management were aware that the third-party
product would be localized and shipped worldwidebut initially, many
issues regarding localization were not clearly identified. These included:

Is the product internationalized? Is it localized?

Is a localization kit available?

134 Making a Global Product at MapInfo Corporation

Are documentation resources for book, tutorial, and online help

available?

Does the product support all character sets, including double-byte

languages such as Japanese and Chinese?

Are bugs directly related to localization easily identified and fixed?

Fortunately, midway through the project, our localization department was

able to open communication lines with the third party and identify and
resolve these problems.
MapInfos Third-Party Strategy

MapInfo has created a Third Party Technical Group (TPTG), which is part
of QA, to address these scenarios. The TPTG joins the MapInfo/thirdparty discussions at the contract-negotiation stage. The TPTG submits a list
of questions to the third party to learn about their development practices,
differences in terminology, etc.
Once development begins, the TPTG is responsible for monitoring the
progress of the project and making sure that designated milestones are met.
MapInfo TPTG conducts basic stability testing of the third-party deliverables and provides feedback as needed. If at any stage MapInfo TPTG feels
that the proposed deadlines may not be met, it notifies the involved parties
and a joint decision is made on how to proceed. This kind of technical
involvement as early as the contract-negotiations stage helps MapInfo and
the third party develop joint products in a shorter time frame while maintaining quality.
Quality Assurance/Worldwide Beta Prog ram

For our English products, MapInfo has an extensive beta program that
includes some international partners. Beta programs are crucial in helping
QA ensure that a product meets corporate quality-assurance standards.
For each international release, we incorporate a similar program on a much
smaller scale. For obvious reasons, we cannot rely on the company doing
the localization to also perform QA. Our strategy is to enlist strong users in
each country to participate in the beta program. We usually allow a few
weeks of extensive testing and then collect feedback from the beta sites.
This information is then analyzed to determine if any problems are localization- or core-code-related, and take appropriate corrective action.

Ricky P. Thibodeau 135

Worldwide Strateg y
Schedules

If MapInfo is to be successful overseas, our international release schedule

must be coordinated with our overseas offices. The localization department
prepares a detailed analysis to determine when localization kits can be
assembled, including availability of documentation resources. We then
determine an overall localization schedule, including the timeframe for
actual deliverables of translated resources from the localizers. All of this
information is communicated to our UK and Australian offices to ensure
that schedules can be met according to quarterly product-delivery goals.
Vendor Selection

The localization department works closely with the operations manager in

the UK and with our corporate Asia/Pacific team to determine which vendors and distributors will be used for the next localization project. Our plan
is simple: select a distributor/SLV that we have either used in the past and
know can deliver quality products on time, or follow recommendations
from our industry peers. Once we have agreed on a localizer, we set up an
agreement, establish communications (via Lotus Notes), and conduct
proper training. For first-time localizers, we usually visit the site to make
sure proper resources are available, and also to conduct detailed training on
localizing our products.
Country Managers Role

Our international country managers are essential to maintaining a strong

distribution network. They also have the task of creating a business plan to
support the introduction of new languages to be localized. Contrary to
popular belief, at least at MapInfo, the localization department does not
decide which market to penetrate next. Rather, the country managers
business plan helps identify the potential profitability for that untapped market. Cost/benefit analysis needs to encompass any engineering and localization costs incurred by introducing a new language.
Producti on Issues

MapInfos general production manager works with the manager of each

product to determine the bill of material (BOM). Every product is issued a
unique part number, and each component of the BOM is considered to be
raw material. The raw materials are each issued a part number that may be
included in many products. Examples of raw materials would be the program CD, manuals, serial-number stickers, registration cards, product boxes,
and license agreements. By keeping a generic registration card or product

136 Making a Global Product at MapInfo Corporation

box that is suitable for multiple products, we can order a higher volume at
lower costs. In our accounting system we take the orderable item-part number or parent number and build the raw material part numbers underneath it, thus creating the BOM. The accounting system takes all costs
associated with a specific raw material and creates an average cost, which is
then rolled up into the cost of goods sold for the parent number.
Determining print quantities is a particularly difficult task for a production
manager. When a new manual is being written, the product manager must
work with the international offices to forecast the number of manuals
needed for the current quarter and anticipate for subsequent quarters. From
corporate headquarters in Troy, New York, our production department
ships directly to all customers in the US and Canada. Internationally, we
ship to our foreign MapInfo offices, distributors, and resellers.
Our previous model required all production and fulfillment to be done
from our corporate headquarters. With our current production model, production (media duplication, printing) and fulfillment for all European languages is handled in the UK. Production for Japanese and Korean is
currently done in the United States, Simplified and Traditional Chinese in
Beijing. Our future plan is to centralize all Asia/Pacific languages in Asia.
MapI n fo L o c al iz a t io n D evel o pm e nt C yc l e

To achieve simultaneous language releases, a company must develop localized versions in sync with original product development. This is often
unfeasible, particularly for smaller companies, due to the additional
resources required. MapInfos goal is to consistently release the major languages as close to the English release as possible. In the past, for a major
point release with new documentation, this could take eight to 12 weeks.
Smaller maintenance releases without documentation rewrites would usually occur six to eight weeks after English shipments. We have instituted a
strategy in which we use the product-development schedule as our main
tool for producing deliverables. We call this the localization-development
cycle. We can now ship major releases in six to eight weeks, and maintenance releases in four to six weeks.
Defining the Process

Product development is a documented process with specific schedule

requirements and associated target timeframes. The outline gives us an
overview of the entire localization process, in a form that is easily communicated to our international offices and localizers. To get a head start on UI
localization, we release a beta localization kit three weeks before UI freeze.

Ricky P. Thibodeau 137

We then give the localizers the delta kit a few days later. By the time localizers receive documentation resources, the UI component and initial builds
of the product should be complete and testing should be under way. Below
is a schematic of our localization development cycle.
12-16 weeks
3 weeks
Design
Finalized

Beta

Delta
Loc Kit

4 weeks
UI
Freeze

2 weeks
FCC
Build

3 weeks
Docs
Freeze

4-6 weeks
English
SHIP

INTL
SHIP

Build Loc Kit

(Delta, Delta)
Post to Loc Sys
Database

Test

Localizer
download & tests
(local QA)
SETUP BETA
PROGRAM

Send to Localizers

L10N (UI,
Documentation)

Test (compile
resource DLL's)

NO

OK?

NO

Build Master CD
OK?
Send CD to
Localizer

Send resources to
MapInfo

Build Localized
Versions

Localizers Final
Test
NO

OK?
Internal Tests
SHIP!

To o l s : L o t u s N o t e s a n d t h e I n t er n e t

Thanks to Lotus Notes, we now have a communications model that enables

us to distribute localization-related resources (localization kits, installable
images, etc.) to localizers worldwide. Notes enables the team to download
localization files and updates, and saves us weeks compared with our previous model of shipping resources on diskette. In the past (1995), the only
communication between localizers and Corporate was fax or mail. Now we
can have live discussions and respond to problems the same day, as the
following illustration shows.

138 Making a Global Product at MapInfo Corporation

I n tte
e rn e t
IIS
SP
P

F ir e w a ll

L o tu s N o te s
L o c a l i z a ti o n
A p p l i c a ti o n

M a p In fo
P r o d u c ts
( E n g lis h
B u ild )

I n te r n a tio n a liz a tio n

( M u lt i- B y t e
E n a b le d )

L o c a l i z a t io n

Since 1995, we have been using Notes as our main source of communication with our worldwide distributors and localization vendors. If they have
access to the Internet, our systems can be configured to allow them access
to our Localization System and Discussion databases (described below).
Notes was selected as our main communication tool with localizers for
these reasons:

It is a document database system that communicates over the Internet.

It allows database replication to client machines (important because our

localizers cannot be expected to install servers).

Since the database is large, scheduled replication can happen overnight,

while Web pages must be downloaded manually, document by
document.

Localization System Database

A Notes database was specifically designed for posting our installable image
files and localization-kit resources. With access to our databases, a localizer
can download the latest localization kits and start work immediately. Also,
once we have built a specific language version for them, they can download
the images immediately after we post them. They can then conduct internal

Ricky P. Thibodeau 139

testing and make changes as needed. This process continues until we are
ready to produce a CD master.

Our Notes administrator is responsible for content and assigning access and
security rights. For example, when we post the Japanese MapInfo Professional installable images, only MapInfo Corporate and our Japanese localizer would see this posting.
The following screen shows the layout for posting files to the database. We
determine the options for the language, product, file type, version, and
build number, as well as a checkbox if the software is copy-protected.

The following figure shows the layout of the localization-kit posting screen.
This is the view localizers have when they are ready to download specific
localization resources.

140 Making a Global Product at MapInfo Corporation

Localization Discussion Database

A discussion database was also developed which enables open communication among localizers throughout the world. Due to our relatively small
staff, this database became critical in resolving issues. To be an effective tool,
the database must be monitored daily and localizers encouraged to participate. Someone must be designated as the database owner to ensure active
participation.

One can display topics by category, author, or favorites. Access and security
rights for each database are set at the field level. With the discussion database, we grant localizers rights to post and edit their own documents and
respond to all documents.
Below is an example of the thread of communication, and how localizers
can help each other resolve problems. In this example, a Chinese localizer

Ricky P. Thibodeau 141

posts a problem communicating with other localizers. A Russian localizer

responds with a solution.

Quantitative Analysis

A quantitative analysis of each localization model requires a separate examination of costs associated with each language and then an overall compilation for each model. Each table shows the cost of the model as a percentage
of actual translation cost of the full localization budget. First is a breakdown
prior to 1996:
Model
Distributor
SLV

% of Actual
% of Full
Translation Cost Localization Budget

Languages
12

55%

30%

0%

0%

MLV

1 (Japanese)

45%

25%

Total

13

100%

55%

Analysis: Fifty-five percent of the full localization budget was used for direct
localization cost associated with translating 13 languages. Twenty-five percent of that budget was for one language. The actual cost of the single language done by an MLV was about 80 percent of the entire amount for the
12 languages done by the distributors. For 1996 to present, the breakdown
is as follows:

142 Making a Global Product at MapInfo Corporation

Model
Distributor*
SLV

% of Actual
% of Full
Translation Cost Localization Budget

Languages
14

60%

40%

40%

25%

MLV

0%

0%

Total

20

100%

65%

*Documentation resources outsourced for three of these languages.

Analysis: Sixty-five percent of the localization budget was used for direct
localization cost associated with translating 20 languages. There is much
better balance in our current model. The actual cost of the six languages
done by the SLVs was about 60 percent of the entire amount for the 14 languages done by the distributors. Cost per language as compared to the previous model was reduced by 25 percent.
Note that had we not used distributors in our model, the cost of exclusively
using the SLV/MLV model would have been 55 percent more for actual
translation expenses, a savings of over US$300,000. Additional savings have
been realized by implementing Lotus Notes our primary mode of data
transmission.
I nt er nat i o na li z a ti o n/ L o c a li z a ti o n P ro ble ms an d I s s ue s

Updates: Trying to release the international versions shortly after the English
version requires us to provide several localization-kit updates to our localizers. In the past we have sent updates to the localizers as soon as we received
them. Depending on the stability of the English software, this could amount
to several updates. Now we gather them and send them as needed. Our
goal is to release a delta localization kit usually three weeks before UI freeze
and then release a final delta one week after UI freeze, resulting in minimal
disruption to the localization process. We maintain a public database with
all the changes and the associated files.
Bug fixing: After release of a product, bugs will inevitably be found. Whenever we get bug reports from our localizers, the localization department is
responsible for duplicating the bug and determining if it is localizationrelated or also part of the English build. There are now plans to start posting
corrections for international builds on our US Web site.
Schedule nonadherence: At the beginning of the localization process we set up
a schedule and then check with the localizers to see if they have the
resources to meet this schedule. Generally, they agree and sign a localization

Ricky P. Thibodeau 143

agreement. If not, we work around them to find another satisfactory

release date.
Ongoing scheduling problems: Once the schedule is agreed, we still have localizers that do not adhere to dates for resource deliverables. Since distributors
are performing localization for sales credits, their sense of urgency is not
high. We have moved large-market languages to SLVs where there is much
higher success in getting projects done on time.
QA issues: We have initiated an international beta-testing procedure for
localized versions. We plan on investigating public betas for international
versions.
QA testing process: The small size of the localization department coupled
with so many language-specific releases forces MapInfo to rely heavily on
outside localizers and distributors to perform most testing. However, as
products become more complex, localization bugs are being discovered
more frequently, requiring localization engineers to spend a significant
amount of time manually testing the product. With 20 language versions,
this can be extremely time-consuming and is often unreliable. In response
to this, a prototype Localization Automated Testing System (LATS) has
been developed internally. It is designed to maximize efficiency and reliability and minimize manual testing. LATS test script can work with all 20
languages that we support. There are pros and cons for writing universal test
scripts, as outlined below:
Pros:

Single test script can run on all 20 languages, therefore it does not
require us to rewrite for each language.

When we modify the test script due to a feature and/or version change,
we need only update one test script.

Cons:

Since test scripts must work on all 20 languages, you can only test very
common operations. Therefore, LATS is a very powerful tool for
smoke testing.

Writing a universal test code is very tricky. A condition may work for
some language but not for others. The test script must be fully tested to
make sure it will work in all 20 language environments, which can be
very time-consuming.

144 Making a Global Product at MapInfo Corporation

Hardware issues: We used to keep multiple operating systems on a single

machine but now we create images that allow us to clean the machine
before doing installations. We create these for about 16 languages under
Windows NT.
Estimating project-localization time: We estimate the time we think that it
should take someone to localize any given project by doing a detailed analysis of the delta localization kit as well as the full localization kit, usually
based on number of words and number of files, dialogs, etc. required for
modification. This estimate works for about half of the localizers. It seems
that skills and productivity vary greatly. We generally use one of our better
SLVs quotations as a benchmark.
Inter nati ona li za ti on ( I1 8 N) and L oca l iza tion ( L1 0N ) Proble ms

Below are some typical problems we have encountered.

I18N Problems:

Third-party products: In a 3D graphics dialog, items in the Type field

shown below as Tyyppi are hard-coded in English. This may be due to an
internationalization error, or might be localization-related if localizers forget to translate the strings.

In the MapInfo Preference dialog, there is no scroll bar in the directory

list. Therefore some items are not accessible using the mouse. In the English
dialog, all items are shown. However, some languages require more spaces
to show all items. This can be remedied in two ways.
1. Localizers can adjust the list to show all items. This method requires
localizers to make this adjustment each time there is a new version,
since we usually add more items in each version. This is our current
approach.
2. If development adds the scroll bar, then localizers need not adjust the
dialog each time. We hope to move to this solution.

Ricky P. Thibodeau 145

L10n Problems:

Localizers do not always update the resources we send them. This shows up
in many places such as a missing Installer string in the Russian example
below.

Localizers must convert any characters above an ASCII value of 128 into
their own character or use some other representations. For example, half
can be written as in Latin1. Internally this representation is 189 in the
Windows Latin1 code page. But in Latin2, like Polish, 189 is the double
acute accent illustrated in the dialog below. There is no in Latin2, and
localizers will need an alternate representation such as 0,5.

Su mm ar y

In this case study, we have attempted to show the importance of a centralized localization strategy. Even a smaller company attempting to tackle
international markets can, with some ingenuity and discipline, develop the

146 Making a Global Product at MapInfo Corporation

systems and procedures needed to deal with localization and translation in a

proactive, strategic manner.
We have also found that the latest technology, most notably the Internet and
workgroup solutions such as Lotus Notes, can prove invaluable in speeding
the localization process and ultimately time-to-market.
Finally, we at MapInfo have taken important first steps in attempting to
develop quantitative models for gauging price and quality. Though much
remains to be done, as of the present writing, we have turned localization
and internationalization to a largely predictable process with predictable
and profitableoutcomes.

10

L A N G U A G E TO O L S
C A S E S T U DY

Section IV: Language Tools and Techniques

10

A Quality-Assurance Model
for Language Projects
Siu Ling Koo
L&L Gebruikersinformatie B.V.
Harold Kinds
Van Zwol Wijntjes

Translation quality is difficult to judge, let alone quantify, and the question
of how to measure it has long been subject to debate. Not only are conflicts
between clients and translators common, but translators themselves also regularly fail to agree on basic principles. What is needed is a rigorous yet realworld approach, one that can be used easily, reliably, and consistently, often
under intense time pressure.
This article documents an attempt to meet these requirements. L&L, a
Netherlands-based translation and localization services provider, has applied
the Quality Assurance Model of the Localisation Industry Standards Association (LISA), and a translation sampling method developed for use with it.
The tools used to grade both translators and translations have led to a more
explicit, objective, and uniform understanding of quality among translators
and reviewers. In addition, early sampling has allowed early feedback and
error rectificationcritical in the fast-moving localization business.
T H E

L I S A

Q A

M O D E L

The QA Model used was developed by a LISA Special Interest Group

(SIG). LISA, a nonprofit association, is a global trade organization for localization, internationalization, and multilingual publishing. In addition to
holding four forums a year throughout the world and publishing a regular
newsletter and other industry information, LISA promotes the development of processes, tools, and technologies that support the localization
business.
The first version of the LISA QA Model was published in 1995, and grew
out of real-life quality metrics provided by companies such as Microsoft,
Digital Equipment, Rank Xerox, IDOC Europe, DLS, and IBM. Work on
version 1.0 was coordinated by Teddy Bengtsson of Oracle. The metrics
were adapted to produce checklists for formatting QA, functional QA, and

148 A Quality-Assurance Model for Language Projects

language QA, along with a reference manual and QA form templates. Version 2.0, released in August 1999, has been expanded to cover all aspects of
localization, from documentation, help, and software to packaging and
CBT (computer-based training) tutorials, as well as including Asian-language support and a Microsoft case study. Companies involved in the work
included Bellcore Learning Support, Hieronymous, International Communications, Language Management International, Microsoft, Nokia, One
World, Oracle, Prolangs, SDL, Augur, and Compuware, with the coordinator being Siu Ling Koo of L&L.
The basic ideas driving the QA Models approach are repeatability (one person doing the same work twice should obtain the same result) and
reproducibility (two people doing the same work should also obtain the same
result). In addition, the Model makes a fundamental distinction between
quality assurance (QA), with checks based on sampling, and quality control
(QC), which entails a 100-percent review. Thus, in the LISA QA Model,
quality assurance is performed twice during projects: on the translations
when first submitted, and on the final product (i.e., following QC and
implementation of the resulting changes):
The LISA QA Model Workflow
QA Translated Text
QA Template

Pass?

Yes

Full review (QC)

Corrected file
instructions for
improvement

QA Reviewed Text
QA Template

No

Add instructions for

improvement in
Comments field

Siu Ling Koo & Harold Kinds 149

In sampling, texts corresponding to a predefined percentage of the total

volume are collected and tested for specific categories of errors. The share
of texts selected is often 10 percent, although the figure can be customized
to individual needs.
L&L has chosen to enhance the LISA QA Model with more rigorous sampling methodology, and has developed its own statistically based sampling
approach. Errors are broken down into a number of categories and classified
according to severity. They are recorded on special Excel templates supplied
with the Model (see figures below).
The presence of even one critical error means that the work is rejected (i.e.,
returned to the supplier for rework with details of the error); otherwise,
major and minor errors are weighted and counted towards a maximum permissible number. This total is then translated into a final Pass or Fail for the
sample (and, where applicable, for the entire job).
P A S S

O R

F A I L

C R I T E R I A

Error levels are defined as follows:

Cri t ica l e rror

an error in a particularly visible part of a document or software

(first screen, document title, etc.);
a major error repeated in several locations;
a localization error causing an application to crash, negatively
modify/misrepresent functionality of the software;
weighting: maximum number of error points allowed plus one (+1).

Ma jor e rror

an error in a highly visible part of a document or software

(header, TOC, menu command, help/CBT topic);
an error resulting in a misleading statement;
a minor error repeated in several locations;
a case where previous QA feedback was not applied;
an error resulting in a potentially offensive statement;
weighting: five error points.

Mi nor e rror

any error of a lesser severity than major;

weight: one error point.

150 A Quality-Assurance Model for Language Projects

The default for the maximum number of error points allowed for the sample to pass is set at one percent of the total number of words in the sample.
However, the client and/or service provider may customize the maximum
error points allowed, how these allowed points are distributed among the
error categories, how the error types are defined, and how the templates
themselves are configured. This ensures the necessary flexibility across different types of projects, languages, and cultural environments.
LISA QA Model Error Categories
Punctuation
Spelling
Style
General style
Register/tone
Language variants/spelling
Country
Country-specific standards
Local suitability
Company standards
Consistency

Mistranslation
Accuracy
Omissions
Additions
Cross-references
Headers/footers
Terminology
Glossary adherence
Context
Language
Grammar
Semantics

Sample LISA QA Model Template

U S E

O F

T H E

L I S A

Q A

M O D E L

AT

L & L

L&L mainly works for businesses in the IT industry. The firm provides
translation, desktop publishing, engineering services, and localization, the
adjustment of a product to cultural and local circumstances.
Time pressure is always high and the need for reliable quality vital. These
circumstances have made it necessary to develop a system that speeds up
review and delivers a level of quality that can be guaranteed within predict-

Siu Ling Koo & Harold Kinds 151

able limits. Two years ago, the company introduced the LISA QA Model,
developing its own sampling methodology to streamline the review process.
With sampling, a translation can be evaluated quickly and the level of confidence that the acceptable quality level has been achieved is known. Beyond
cutting review time, sampling also results in a classification of the errors
found by the reviewer along with error trends. These can help the producer
prevent repetition of these errors in their implementation of corrections
and in future jobs.
L&L found it rather difficult at first to get reviewers to work with the
model, especially those used to doing complete reviews. It is of great
importance that taste is not an issue and that only those errors that are justifiable are actually recorded. Introducing such subjective judgements renders
a sampling method of little use or difficult to implement. However, experience showed that after reviewers had used the model for some time, they
came to appreciate it. In fact, in judging a text they preferred to have the
quantification that the model provides them, something they could not
achieve without this approach, even if they were to read the entire text
rather than a sample.
One of the problems often mentioned by reviewers was that it was very difficult to distinguish between a major and a critical error, or a major and a
minor error. For the model to work properly, critical errors must be carefully defined before a project starts. To avoid confusion, errors should also
be defined similarly for projects involving similar types of materials. For
example, errors in titles, phone numbers, addresses, and index entries in
user manuals should be consistently defined as critical errors. The difficulty
in distinguishing between major and minor errors is more common and
more difficult to resolve. The choice of defining an error as major or minor
may become significant in the overall evaluation, and only consensus
between the producer and the receiver will resolve to which category a particular type of error belongs.
Furthermore, although the LISA QA Model is based on sampling, a reliable
sampling method is not available for it at the moment. Reviews are mostly
performed on samples selected at random by the reviewer. After a quick
browse through the materials, the reviewer will generally have a feeling of
what to select for the sample, based on which section is the most difficult or
representative of the whole. However, in order to have a truly reliable QA
model, a more systematic sampling method must be developed. With a reliable sampling method, a higher confidence level in the accuracy of the Pass
or Fail score, and ultimately in the quality of the final product, can be

152 A Quality-Assurance Model for Language Projects

achieved. For this reason, L&L has augmented the LISA Model with its
own sampling methodology, outlined below.
L & L

S A M P L I N G

M E T H O D O L O G Y

I ll u s t r a ti o n o f t h e S am p l in g Con c e p t

The idea behind sampling can be illustrated as follows:

A manufacturer of bolts wants to guarantee the quality of the bolts he sells.
This means that he has to perform a quality check. He has two options:
1. check all bolts that are produced; or
2. devise a plan with which he can guarantee to a certain extent that his
bolts will pass his customers requirements.
Option 1 is clearly impractical, expensive, and time-consuming.
Option 2 requires a sampling plan. He must first agree with his customer on
the number of bolts that can fail and then determine his confidence level
that the bolts he delivers comply with his customers requirements. In other
words, he must determine how large the sample should be and which particular bolts are to be quality-tested.
If the bolts are to be delivered in boxes (clusters of bolts) he can also determine which boxes to check and then proceed accordingly.
Methodolog y Developed at L&L

In a similar manner, a tool was developed by Harold Kinds, statistical consultant to L&L, to determine the acceptability of a text based on a sampling
plan in conjunction with the error definitions, weights, and criteria as outlined in the LISA QA Model.
A sampling plan should consist of these two components:
1. determination of a sample size;
2. determination of a procedure to select a sample.
These steps can be approached in a variety of ways. We chose to base our
sampling plan on a technique for monetary-unit sampling.1 At L&L this
sampling plan has been in use for a several months, and the method will
continue to be monitored and adjusted as necessary. A brief overview of the
method and the results to date are described below.
1. J.H.G. Kinds, Virtual Sampling Techniques for Monetary Unit Sampling in Statistical
Auditing, Statistical Auditing Conference, The Hague, 1995.

Siu Ling Koo & Harold Kinds 153

Sampling Plan

The L&L sampling plan is derived as follows:

1. The translated text to be reviewed consists of x words.
2. Each word has the same importance, and all words can be regarded as
equivalent units. This is not strictly true in actual language usage, but it
is nearly impossible, if not impractical, to assign varying importance to
word categories.
3. The sample will be based on the original document, which has a oneto-one relationship with the translated text. Again, this is not
completely true for each individual word but will be true for the
selected group of words.
Sample Size

Sample size is determined by the level of confidence with which the producer can assume that an acceptable error limit will not be exceeded. Both
the confidence level and the acceptable error limit are set at the beginning
of the project by customer and producer. (Determination of the size is
based on Poisson distribution as an acceptable, even somewhat conservative,
approximation of the binomial sampling distribution.) If the level of confidence and the acceptable error limit are changed, the size of the sample
changes accordingly.
For example, if the level of confidence is set at 95 percent and the acceptable error limit (inaccuracy) for accepting the translation is set at 2.5 percent, this would lead to a sample size of 120 words.
Real Sampling Design

After determining the number of words in the sample, the question is how
to select these words from the text in such a way that the reviewer can have
a reasonable opinion about the quality of the translated word. In translation,
it is impossible to check on correctness per word. This would mean that a
sample of a number of random, individual words would be taken and
checked as to their correct translation, as was done in the illustration with
the bolts.
However, a word always occurs in a context, and cannot be checked on its
own. A statement on the quality of the document arrived at by checking
120 separate words would therefore be meaningless.

154 A Quality-Assurance Model for Language Projects

Virtual Sampling

Virtual sampling does not look at random words, as would be the case in a
real sampling design, but rather at clusters of words. Based on the same
probability distribution, we can compute the probability that one or more
words of a particular cluster are within the selected words of the sample. If
one or more words in the cluster are selected, this cluster must be reviewed.
The size of the cluster is not important and can vary within the project.
The words are not checked individually but are in the context of a page, for
example. The number of words on each page is counted and regarded as the
cluster. Since clusters can vary in size, pages can be used as a means of clustering. The same goes for paragraphs, or other units that seem reasonable.
Virtual sampling can be refined to more inspection levels, which can be
regarded as different sampling plans. If a page contains one or more selected
words, it will be inspected for at least one or more critical errors. The probability can now be determined that one major error which, according to
the LISA Model, is the equivalent of five minor errors, or a minor error,
which is equivalent to one word, will be selected. The determination of
this probability is computed using the same method as described above. For
each page the probability is determined that:

a critical error occurs on a page, which results in a minor inspection of

that page;
a major error occurs on a page, which results in an average inspection
of that page;
a minor error occurs on a page, which results in a tight inspection of
that page.

Example

A sampling plan for a 21-page document based on a 95-percent confidence

level and an inaccuracy level of 2.5 percent appears as follows:
Page

No. of Critical
Words Error

Major
Error

Minor
Error

Inspection Plan

12

no inspection

120

minor inspection on critical errors

129

minor inspection on critical errors

104

minor inspection on critical errors

43

minor inspection on critical errors

228

average inspection on major errors

Siu Ling Koo & Harold Kinds 155

Page

No. of Critical
Words Error

Major
Error

Minor
Error

Inspection Plan

180

minor inspection on critical errors

199

minor inspection on critical errors

222

minor inspection on critical errors

10

96

average inspection on major errors

11

165

minor inspection on critical errors

12

171

minor inspection on critical errors

13

53

minor inspection on critical errors

14

168

minor inspection on critical errors

15

23

minor inspection on critical errors

16

114

average inspection on major errors

17

214

minor inspection on critical errors

18

124

tight inspection

19

236

minor inspection on critical errors

20

37

no inspection

21

no inspection

In this example, 18 out of 21 pages, or 86 percent of all pages, must be

inspected for critical errors (entailing minor inspections). This means that
these 18 pages must be scanned quickly, since critical errors are mostly very
obvious errors. If a critical error is found, the document is rejected and
returned to the producer.
If no critical errors are found, the document is inspected for major errors.
In this particular example 3 out of 21 pages, or 14 percent of all pages,
require average inspection. Thus, the pages where major errors may occur
are reviewed with more care than in a minor inspection. The pages
inspected for major errors have already been inspected for critical errors.
Thus, a page requiring an average inspection has always been inspected at a
minor inspection level. If a major error is found, the document is rejected.
If no major errors are found, the document is inspected for minor errors. In
this example, only one of 21 pages (ca. five percent) requires tight inspection. A tight inspection always includes a major and a minor inspection. A
tight inspection means that the page where the minor error may occur is
subjected to careful and detailed reviews. If a minor error is found, the document is rejected. If no minor error is found, the document is accepted.

156 A Quality-Assurance Model for Language Projects

Practical Experience

The following is an outline of actual results when this system was applied.
L&L was contracted to review translations for a particular customer. The
contract duration was for one year and translations produced by several vendors (producers) were to be reviewed. Turnaround time was critical. It
was of great importance that translation quality was assessed quickly so that
decisions could be made as rapidly as possible as to whether or not to send a
problematic translation back to the producer or have the reviewer implement corrections.
The first problem encountered was that the customer wanted quality but
then failed to clearly define what quality meant to them. In other words,
definitions for critical, major, and minor errors could not be given. In this
case, L&L used its own standards and hoped that this definition suited the
customer as well. At the beginning of each review cycle, certain information was routinely checked independent of the sample. This included article
numbers, phone numbers, addresses, etc. If such information was incorrect,
a Fail (according to the LISA QA Model) was given and the producer
received a warning. If the review of the sample received a Fail, the translation was returned to the producer.
L&L is using the sampling model in two stages of translation. First, it is used
to sample translations at the start of a project. After about 5,000 words have
been translated, the text is reviewed to assess whether or not instructions
given at the start of the project have been followed. On pages where a tight
inspection has to be performed, errors are categorized to give feedback to
the translator. The following categories are used:
Grammatical and typing quality
Mistranslation
Style
Country-specific standards
If necessary, instructions are given to improve the translation. In extreme
cases, translators have been pulled off the job. This has proven to save time
for the reviewer. When several translators work on the same project, this
check helps to establish consistency, apart from the other advantages mentioned earlier.
Sampling is again employed when review of the total project starts. All
words per page are counted and a sample is taken. The levels of inspection
per page in the sample are computed and review starts accordingly. It will
be necessary to randomly perform a complete review to see if the method
remains valid under varying circumstances.

Siu Ling Koo & Harold Kinds 157

N E X T

S T E P S

L&L started using the sampling method recently. In the beginning, projects
that had already been reviewed were tested against the results of a traditional
review. The results were very encouraging. The next stage is to use the
sampling plan and then use the traditional method to compare results. This
will be done for several months, in order to have confirmation that the
method is valid and for the reviewers to gain confidence in the method.
After this stage has been satisfactorily completed, customers will be
introduced to the method. The model can then be customized for
each customer.
The time in which a review is performed can be greatly improved by using
valid sampling methods and valid quantifiable criteria to check the sample.
Greater certainty about the level of the quality received will also be a natural result. The application of quantitative methods to the language business
will, we are confident, greatly help project throughput and quality.
Part of this case study is based on an article by Deborah Fry in Language International, October 1999.

11

L A N G U A G E TO O L S
C A S E S T U DY

11

Terminology Management at Ericsson

Gary Jaekel
Ericsson Language Services

A glance at any business newspaper reveals the impressive strides in telecom

technology in recent years. Behind the headlines, a silent revolution has
accompanied these developments: consumers are placing ever more stringent demands on the quality of the information that accompanies these
products. Further, as sluggish national monopolies have been displaced by
private telecom operators, the new ventures can ill afford downtime and
delays in time-to-market. Telecoms are increasingly dependent on product
information that is clear, concise, and accurateboth for subscribers and
internal staff. Miscommunication is something providers feel on the bottom
line.
One key to clear communication is effective management of terminology.
Terminology databases can help ensure greater consistency in documentation, which can not only make documentation easier to read and understand, but also avoid miscommunications that might result in customer
dissatisfaction or give rise to potential product liability.
Recent years have also seen the rise of more sophisticated tools in managing
terminology, including translation memory, a technology that allows users
to leverage or reuse existing translations. Effective terminology management cuts costs, improves consistency and hence linguistic quality, and
most important in this era of intense time-to-market pressuresreduces
turnaround times for translation.
In this case study, we examine the implementation, over a span of years, of
terminology management at telecom giant Ericsson.
B A C K G R O U N D

Ericssons commitment to terminology management is hardly a recent

development. From about 1980 to 1985, a group of translators at Ericsson
worked to publish a telecommunications glossary in five languages called
Eriterm. Eriterm was published in book form in 1985, and featured 13,000
Swedish terms and their equivalents in English, French, and Spanish (the
original goal of including German was not met at the time). Each of some
13,000 term records contained a domain, but no definitions. Eriterm grad-

160 Terminology Management at Ericsson

ually became well established within the Ericsson world, although it was
not afforded the status of a standard. At the same time, a database called
Term resided on a mainframe, and was administered by what was then the
Standards Department. This database included most of Eriterm, as well as
many other terms submitted by various units within Ericsson at their own
discretion.
In the late 80s under my management a new project set out to republish
Eriterm in five volumes: each proceeding from one of the five languages,
this time with the inclusion of German. Around this time, responsibility for
Term was transferred from the Standards Department to Ericsson Language
Services. Aside from the need to update the material in Eriterm, the idea of
using any language in the original data as a source language created problems of which we were originally unaware. Since the original Eriterm used
Swedish as a source, alternate terms in other languages were given that did
not necessarily result when the records were reversed. We also initiated further development of the database so that we could get print-ready copy
directly. This project continued until the publication of the new Eriterm in
late 1992, which contained about 15,000 terms. Besides a broader array of
terms, the new edition included better domain information, but still lacked
definitions.
Once Ericsson Language Services had been assigned responsibility for the
Term database, we felt that the next step would be to make the database
more accessible to our in-house operations, making it easier to upgrade
future releases of Eriterm in whatever form. To this end, we began looking
at the various terminology-management products on the market, while not
completely disregarding the idea that we might develop our own application in cooperation with Oracle, for example. In the end we opted for Tradoss Multiterm as the tool which most closely met our needs. On this
platform we began development of Termelsa, our in-house database.

Gary Jaekel 161

Sample Trados MultiTerm Record

A new edition of Eriterm in book form has not been released since the
1992 edition, and it is extremely unlikely that any new book versions will
be published in the future. On the other hand, Eriterm-PC was developed
in the mid-nineties, first in DOS and then in Windows. Eriterm PC
includes definitions or context for most term records.
Once the dust had settled after the release of the new Eriterm in 1992, we
began to take a more structured look at how we intended to build up and
maintain our terminology competence, as well as keep our product as upto-date as possible. We found we had to:

clearly define who was responsible for terminology for each language
and who would manage both strategy and operations;

establish routines for database management and input of new terms,

both from customer-based and in-house perspectives;

ensure that all aspects of terminology operations were well

documented;

establish plans for further competence development for all involved;

plan for future development, including resource availability.

162 Terminology Management at Ericsson

Terminology Management at Ericsson: Overview

Stage

Year

Content

Scope

Original terminological
material compiled in
pamphlet form in five
languages

1980

general
telephony
terms

5,000 terms

Eriterm in one volume, from

Swedish, and lacking German

1985

Eriterm

12,500 terms

Eriterm in five volumes from

each of the languages and
greatly restructured

1992

Eriterm
vol. 1-5

13,500 terms

First DOS, then Windows

version, including definitions
Reorganized domain
structure

1995-1998

Eriterm PC

14,000 terms

On CD for PC and Unix

with considerably updated
terminological content

1999

Techniterm

16,500 terms

R E S P O N S I B I L I T Y

A N D

M A N A G E M E N T

As we had five languages in which to maintain terminology competence,

one person was named terminologist in each: Swedish, English, French,
Spanish, and German. One team member took responsibility for database
management, and another for terminology management as such. Initially
terminology was treated as a separate unit within the organization, but was
later merged with translation operations.
Interestingly, none of those involved in terminology projects at Ericsson
Language Services were terminologists, either by training or previous experience. This tabula rasa allowed us interesting insights into how a terminology operation can be established.
In addition, our approach to terminology was always more practical than
theoretical. And because Ericsson Language Services was a member of the
Swedish Centre for Technical Terminology (TNC), we were entitled to a
number of hours per year for terminology seminars and the like. We also
attended various conferences and seminars organized, for example, by Termnet, and made trips to TNCs counterparts in various countries. We also
maintained collaborative relations with other companies and organizations
with terminology operations resembling our own. We invested in training
in the various tools we used, including database management, translation
memory, and machine translation.

Gary Jaekel 163

The responsibilities and authority of the terminologists were documented

in individual role descriptions.
Rou ti nes

Well-documented work procedures for all involved are absolutely necessary

to maintain an in-house database of value to the user and, in our case, to
ensure that our product would remain up-to-date and attractive to our
customers.
Term Selection

Naturally, terms from new technology were particularly interesting, as were

those which were in vogue with our customers at any given time. As we
have all experienced, certain topics and their accompanying terminology
tend to flow in waves through any large company, even if the flow at first
seems sluggish. Typically, an area such as corporate training generates an
increasing demand for writing and translating documents, as the subject
gains visibility. This results in a universal attempt to develop new ideas and
approaches, and one fine day, without warning, a flood-tide of consciousness breaks over the company, washing up terminology that may be excellent or questionable by turns, but is nevertheless in common use. At
Ericsson Language Services, we realized early on that these beachcomber
terms would not only enrich the database, but would provide the raw material that we could later refine into new releases of Eriterm. In addition,
these terms and their equivalents in other languages would clearly assist
other translators with similar projects, even if the terms had not been formally approved or fully specified.
The Translator/Terminologist

Our first responsibility was to identify and establish work routines for those
who would collect raw terminological inputthe translators. This was no
easy task, considering its hit-and-miss nature, which involved scanning
individual card indexes, hard-copy lists in binders, the odd flat file on our
PCs, and the brains of people whose expertise we trusted. The next step
was to see that new terms were entered into the database with reasonable
regularity. For translators, this involved two jobs in parallel: handling translation projects that came in, and extracting from each project terms they felt
warranted inclusion in Termelsa.
The terminology groupconsisting of the translators, the database administrator and the managermet once a month to review terms added to the
database in the preceding period. The group extracted the terms deemed to

164 Terminology Management at Ericsson

be of interest, and tried to find equivalents in the languages not already represented, as well as applicable definitions or contexts.
The database administrator would compile a log of terms that had been
queried in the database in that month (a function available in Multiterm).
This log was then examined for terms of particular interest, a judgment
based partially on the number of searches related to it. Any terms that met
the criteria were handled according to the same routines as above.
Any term records that included at least two languages other than the
source-term language as well as a definition or context would then be
assigned a domain, and possibly subdomain, according to an established list.
The term record could then be formally approved and the necessary administrative data assigned.
At the outset of large translation projects, terminology lists were compiled
and used as input to the database. We were further able to apply our inside
knowledge of new product developments at Ericsson and the telecom sector in general, making our terminology decisions particularly informed and
relevant.
Each translator/terminologist was expected to devote four hours a week to
this work, not counting the monthly terminology group meetings, which
might last half a day.
The Database Administrator

Besides the tasks mentioned above, the database administrator was expected
to follow tool and software developments, manage a test database, maintain
a log of any database modifications, keep all documentation current, and
ensure that data were always backed up. As the person with the most indepth knowledge of Multiterm and Termelsa, the database administrator
was also very influential in troubleshooting the overall system and devising
more efficient work methods.
Term Record Classification and Access Authority

Multiterm has eight levels of term-record classification, of which we used

three.

Class 1. This is the standard classification for new term records. All
records receive this designation before they can be developed further.

Class 2. This designation is used once a terms content and English

equivalent are approved. Equivalents in other languages that have not

Gary Jaekel 165

been verified are labeled to be verified under the attribute

Acceptability.

Class 3. When a term record is approved by the terminology group, it is

fixed at classification 3, which means it may be altered only by the
database administrator. All classification-3 terms are also set aside in a
separate database copy for the production of future glossaries.

We have also established four levels of access authority:

Guest. A guest user has read-only access to all term records.

Translator. A translator can input classification-1 term records, but has

read-only access to classifications 2 and 3.

Terminologist. A terminologist can input or update classification-1 and

-2 term records, but has read-only access to classification 3.

Database Administrator. The database administrator can update all three

classifications, fix terms at classification 3, and temporarily upgrade any
users access authority.

Documentation

As an Ericsson company, Ericsson Language Services was required to be

ISO-certified. This meant, among other things, that clear rules governed
the documentation of our processes, products, and routines. Specifications,
descriptions, user manuals and reference guides were drafted, discussed, and
evaluated by everyone working in the terminology team. In particular, it
was important to document how we adapted Multiterm to suit our needs.
This documentation was, of course, updated several times as new ideas and
requirements arose.
Record Str ucture
Domains and Subdomains

In the original Eriterm, we specified such a plethora of domains that we

found the database difficult to administer in the long run. We therefore
decided to group many of the earlier domains into subdomains under more
general major domains. The following is the current domain structure with
a breakdown of term totals.

166 Terminology Management at Ericsson

BUSINESS &
ADMINISTRAT.
business
administration
commerce
dependability
finance
insurance
law
office management
personnel
management
project management
quality
safety
security
standardization
training
DATA
PROCESSING
character recognition
computers
computer graphics
computer security
logic elements
printers
word processing
DESIGN &
DOCUMENTS
design technology
documentation
flowcharts
printing
signs & symbols
R&D
technical drawing

GENERAL
TERMS
MATERIALS
paints and varnishes
plastics
product form
strength
surface properties
MECHANICS
equipment practice
fasteners & threads
mechanical
engineering
MISCELLANEOUS
TECHNOLOGIES
automatic control
buildings &
construction
climate control
environmental
engineering
general tech.
PRODUCTION
maintenance
manufacturing
engineering
marking
metallurgy
product management
soldering
supply & distribution
surface treatment
tools
welding
wire-wrapping
wires & wiring

SCIENCES
chemistry
economics
ergonomics
geography
geometry
geophysics
mathematics
mineralogy
optics
physics
linguistics
meteorology
statistics
TELECOM
AXE
antennas
audio & video
broadband
cables
charging
cryptology
datacom
digital transm
fibre optic com
IN
ISDN
manual switching
military
mobicom
network management
O&M
OSI
outside plant eng
PBX
plant eng
radar
radio

radiowavepropag
space
signalling
switching
switching (code)
switching (crossbar)
switching (rotary)
SDH
telecom networks
telecom services
telephone
setstelephony
traffic engineering
transmission
waveguides
TESTING &
MEASURING
telephonometry

Most terms in the product are related to telecommunications, though they

are likely to be found in other areas of business and technology as well.
Aside from the telecom domain, the obvious growth area is data processing,
which we expect to eventually outgrow and even incorporate telecom to
some extent.

Gary Jaekel 167

New Techniterm: Terms Per Domain and Share of Total

No. of
Terms

Domain
Business & Administration
Design & Documents

% of
Total

730

4.5

353

2.1

Data Processing

2493

15.2

Electricity

1195

7.2

Electronics

981

5.9

General Terms

333

2.0

Materials

279

1.6

Mechanics

466

2.8

Misc. Technologies

683

4.1

Production

734

4.5

Sciences

648

3.9

Telecom

7096

43

327

2.0

Testing & Measuring

Total

16,318

Term Record Structure

The following table shows a term record structure. Though fixed, the structure can be modified based on reasonable user feedback. Not all fields are
required.
Heading

Value

Content/Comment

The following applies to the entire term record:

Creation Date
Created by

Applies to entire record

logged user-id

Change Date

Applies to entire record

Shows the latest change in any field

Changed by

logged user-id

Can be changed at input

Entry class

1, 2, or 3

Shows access authority

Entry number

Number series

[DOMAIN]

Main headings in the Domain List

Subdomain

see Domain List

Subheadings in the Domain List

System

MD110, IDF, etc. To more closely specify a term

Method

PROPS, MQR
etc.

To more closely specify a term

168 Terminology Management at Ericsson

Heading
Glossary

Value

Content/Comment

ELS, ERITERM, Indicates collection of terms the record comes

Ericsson Standard, from
etc.

Definition

English definition is placed before the English

term. Can also contain a translation of a nonEnglish definition

Source

Indicates where a term or definition has been

found (e.g., dictionary, standard, magazine)

Definition note

Connected to the definition or other terms in

the record

Technical note

(very concise)

Technical information about the concept other

than what can be found in the definition syntax

The following headings apply per language:

English

English term or abbreviation

French

French term or abbreviation

etc. for each

language...
Grammar

n; n m; n f; n n; n Additional grammatical information

non-neuter; n pl; Even applies to English and Swedish terms
n mpl; n fpl; n
npl; n non-neuter
pl; n inv; n f, n m;
n m, n n; nf, n; n;
adj; adj inv; adv;
v; vt; vi; vt ind;
phrase

Usage

UK; USA;
Used for regional differences and can also be
AUSTRALIA;
used to indicate that the term is Ericsson-specific
CAN; FRANCE;
BELG; SWITZ;
GERMANY;
AUSTRIA;
SPAIN; LATIN
AMER;
ERICSSON

Source

Where a term or definition has been found (e.g.,

dictionary, standard, magazine)

Customer

ERA, EBC, etc.

Acceptability

correct
avoid
to be verified
preferred

correct=term is approved
avoid=not to be used (an alternative or another
reference is given)
to be verified= term is not fully approved, but can
be used
preferred= term should be used before another
correct term in the record

Gary Jaekel 169

Heading

Value

abbr.

Content/Comment
An abbreviation is written as a normal term
(with an index) and abbr is added (this attribute
has no field value)

Note

(certain standard
formulations may
apply)

1) Indicates limits to the terms use:

- The /French/ term does not apply to
/metalworking/
- applies to keyboards
2) See records:
- See: ^...^ (automatic text link
3) Indicates a variant much like the term:
-Also, logical, for the main term logic
4) Clarifies a term whose context is not clear
from the term itself, e.g.,
<English>equivalent noise voltage
<Note>of a one-port device

Linguistic note
Miscellaneous

Contains information that cannot be put into

another field, e.g., information may later be used
to create a definition or Technical note
Contents not intended for publication

Context

Shows the term in context

Every record also has a Type field, which may contain one of the following values:

A: attribute field

I: index field

P: picklist

S: system field

T: text field

Aside from German terms and proper names, terms are always written in
lower case. Abbreviations are always written in all-caps. The field Linguistic note may be used to indicate that the first letter of the term is normally
a capital letter.
F U T U R E

D E V E L O P M E N T

After operating as an Ericsson unit for many years, and as a company within
Ericsson Group for three years, Ericsson Language Services was bought by
Interverbum in late 1997. The prime purpose of the merger was to access a
larger pool of language resources, and to secure competence in languagerelated technology.

170 Terminology Management at Ericsson

Aside from our established method of working with terminology, Ericsson

Language Services brought with it considerable experience of working with
tools for machine translation. Interverbum, on the other hand, has worked
for several years with a number of translation-memory tools. Because we
are aware that the key to successful use of these tools is a systematic
approach to terminology, we are initiating plans to implement the Termelsa
database as a standard terminology platform throughout Interverbum.
Further, we are examining how we can structure our work with the
database and its content to conform to the needs and routines of a large
organization.
We have decided to change the name Eriterm (due to its strong Ericsson
associations) to Techniterm, reflecting our goal of including terminology
from other areas, embracing, for example, the automotive, IT, and pharmaceutical industries. We have also discontinued the Eriterm-PC product, and
are moving its contents to a new platform, Multiterm Dictionary. We have
found this product easy to use and upgrade, and very cost-effective. Apart
from the new platform, we have two to three thousand new term records
ready for import into the database. The new Techniterm has been slated for
release on CD during the autumn of 1999.
Sample Record from Trados MultiTerm

A Unix interface to Techniterm is under development. Many of our customers, particularly engineers, designers, and documentation professionals,
work in a Unix environment. Work is slated for completion in 1999.
We are constantly being asked when Techniterm will be Web-enabled.
While this presents no major technical problems, we need to carefully
examine how we can be compensated for its use. Several possible solutions

Gary Jaekel 171

are in use in other service or product areas, but further cost-justification is

needed before making this major investment.
Q ua nt if yi ng t he R es u lt s

Below is a breakdown of the income and expenses from terminology management at Ericsson. Monetary figures are in millions of Swedish crowns.
In-House Person-Hours

Cash-Out Development

Sales

9,200

1.6 MSEK

3.8 MSEK

The above figure is based on

hours reported from the
beginning of the 90s, and
does not include marketing
and other similar efforts.

The largest cost here is the

printing of Eriterm in 1992.
Other costs, such as
mainframe database
maintenance, were not
necessarily booked as such
by Ericsson at the time.

Includes the sales of the 1992

Eriterm and licensing
agreements to date. The first
versions of the glossary were
available free within
Ericsson.

Th e R o ck y R o ad a nd t he S up er h igh w ay

While the rocky road of implementing effective terminology management

has now become passably paved, a true terminological superhighway is still
well in the future. Awareness has grown and tools are available, but with
each new positive development, new problems also arise.
One seemingly thorny issue is copyright. A company that pays for development of a glossary pertaining to its area of business is loath to see that terminology used by a competitor. I do not claim to have a ready solution to this
problem, but find it most gratifying that terminology is now considered
something of real value. A similar problem has appeared concerning the
ownership of translation memories. I would maintain that, in the end, the
terms you make are equal to the terms you take.
From my perspective, the greatest problem and the greatest opportunity
before us rest in the transfer of terminology across language-technology
tools. Though this may pose great challenges, the amount of structured terminology in existence as separate islands of competence and reference is
unsupportable in the long run. Admirable work has been done to create
standards, and projects are in progress to facilitate terminology exchange, so
perhaps we are beginning to see the light at the end of the tunnel.
Terminology is gaining visibility as a discipline in its own right, in response
to the heightened importance of user documentation, maintenance manuals, and technical specifications that accompany the vast array of services
and products we encounter in our daily lives.

12

L A N G U A G E TO O L S
C A S E S T U DY

1210

Mission-Critical:
Translating for Regulated Industries
Robert C. Sprung
Harvard Translations, Inc.
I N T R O D U C T I O N

Translation presents acute challenges to companies in regulated industries

those which are subject to domestic or international oversight, and which
must closely tie their processes to larger industrial or governmental standards. Examples include medical-device manufacturers, pharmaceutical
companies, law firms, financial institutions, auto makers, and many businesses where user-safety or legal-liability concerns loom large. The stakes
are high: in the medical-device industry, for example, problems with labeling (which includes translation) are a major cause of recalls.
Much of the translation produced for regulated industries is required by law.
Examples range from the label text found on a medical device or pharmaceutical, to the SEC filing for an initial stock offering, to the emissions
manual for a new car. Many firms are quickly learning that translation is a
subtle task requiring strategic thinking in order to cut cycle time and assure
quality that will pass muster with regulatory authorities. Though translation
usually represents a small proportion of total product-development expense,
glitches in translation quality and turnaround time can create problems out
of all proportion to the cost of such services.
In this case study, we examine the foreign-language solutions developed and
implemented by Harvard Translations over some seven years for two leading
Johnson & Johnson companiesJohnson & Johnson Professional (JJPI) and
Ethicon Endo-Surgery (EES).
JJPI is an old-line medical-device manufacturer. Although a recently created corporate entity, it grew out of Codman and Shurtleff, one of the oldest New England medical-device firms. EES is more entrepreneurial,
specializing in the latest technology for endoscopic surgery. Many EES

1. This case study is an adaptation of presentations, some given jointly with J&J personnel, at
conferences sponsored by Canon Communications and Pharmaceutical and Medical Packaging
News magazine, in which the author publishes a regular column on language issues.

174 Mission-Critical: Translating for Regulated Industries

products have a short life in the marketplace, and every week of translation
time that could have been spared represents significant lost revenues.
Both companies began with little coordinated effort to release their products with labeling in multiple languages. Their main challenges included:

how to organize and proceduralize translation;

how to create a scalable multilingual effort to meet rapidly rising demand;

how to cut translation time to improve time-to-market;

how to measure and control translation quality in the face of a complex

regulatory framework.

The solutions they implemented, together with Harvard Translations, offer

valuable insights for many companieslarge or smallin other highly regulated industries.
C A S E

S T U D Y

1 :

J O H N S O N A N D J O H N S O N P R O F E S S I O N A L ( J J P I )

Ba s e li n e S it u at i o n ( 1 9 92 )

JJPIs situation was similar to that of most large firms exporting in the early
90s: translation was ad-hoc, an afterthought tacked on to the technicalwriting and packaging process.
The Medical Device Directive (MDD) changed all that. In the early 1990s,
this European Community legislation required companies marketing in
Europe to include labeling in local languages (labeling includes all text
associated with a product, whether the printing on the box or the instructional insert inside it). The MDD fueled large-scale translation efforts
many manufacturers erred on the conservative side and translated into languages of countries where they had small current sales, but did not want to
be denied entry in the future (e.g., Finnish and Greek are official languages
of EC countries, and it was not uncommon for even a small medical-device
manufacturer to include these languages).
The MDDand similar legislation in other industriesforced many
companies to take translation seriously, as a strategic issue.
Ch al le nge 1 : T im e -to-M arket

Lacking an organized framework for discussing and analyzing translation,

the language challenge that JJPI first noticed lay in clear sightWhy does
translation take so long? In some cases, it had taken six to eight months to
produce a single instructional insert in seven languages.

Robert C. Sprung 175

The root of the problem was the companys patchwork approach, with
varying sources of translation (some outsourced to local vendors, some
done by individual country distributors) and multiple points of contact
within its own organization for approving translations. J&J had to approve
translations once they were completed, but this approval was not proceduralizedit might be done by a marketing manager in Italy, a regulatoryaffairs expert in Spain, and a packaging engineer in Japan. Many of these
affiliates within J&J performed the approvals in addition to their already
crowded job description. Since this was perceived as an added duty, quality
and timing of approvals were unpredictable.
Translation approvals suffered from what we may call the Rule of the Least
Common Denominator. Since most of these medical devices were designed
for the entire world and required all languages to be complete before shipment,
a single late affiliate reviewer could delay shipment of a product. JJPI
discovered that the dog ate my homework defense was hardly limited to
American schoolchildren: claims of faxes never being received or labels
never sent for approval were not uncommon. Such clerical errors could add
weeks or months to the translation process, causing major shipment delays
and lost revenue.
The company also discovered that affiliates celebrated holidays and took
vacations at different times of the year. When producing a label in 12 languages, one could count on a given day being a bank holiday or saints day
in one affiliate country, or at least one affiliate disappearing for a four-week
vacation. Resolving even a simple question across all languages (e.g., adding
a new caution statement) could easily add two weeks to the process.
Finally, J&J realized it had little political clout over reviewers. An affiliate
might resent marching to the beat of a foreign drummer when reviewing
foreign-language text, and might doubt a US firms ability to speak his or
her language accurately. Though late delivery of a review might have profound repercussions on a products ship date and revenue picture, headquarters had little recourse in such cases.
Sol ut i o n: C en tr al iz a t io n

The first key to reining in the problem was centralizing the process.
Together with this J&J company, we created:

a single set of procedures;

a single hub for project management/trafficking;

a single contact point for assessing project status and coordination.

176 Mission-Critical: Translating for Regulated Industries

The first step in creating a standardized process was to batch translations.

In the past, small installments of labeling text had been sent to translators
and affiliate reviewers as soon as they were ready. We developed a standard
process by which batches of 10,000 English words would be translated and
laid out (using standard desktop-publishing software such as QuarkXPress).
A larger volume of text to work on yielded immediate economies of scale:

It became easier to recruit top translators for highly specialized

linguistic work. The qualified pool of talent is already very small, and
the pool of those willing to drop everything for 500-word assignments
is minuscule. Without batching, top vendors can easily be busy with
other assignments and, even if available, might charge hefty minimum
fees for short translations. Batching enhanced recruitment and retention
of top vendors, which bolstered consistency and quality.

Project-management costs dropped dramatically as a percentage of total

revenue. One could manage 10,000 words in seven languages in about
the same time as one could manage 1,000.

Batching enhanced translation consistency and quality. In the past, J&J

tech writers would send an insert text for translation, and six months
later J&J packaging engineers might send for translation the text that
would appear on the outer box for that same product. The danger
could exist, absent a rigorous process to ensure consistency, for the
Spanish or Japanese product name to appear differently on the outer
box than it did on the instructions inside. Such a translation error
would surely send off alarm bells in the case of an audit, and could
potentially trigger a recall. Batching significantly reduced such risks,
since a batch included most if not all translation related to a product.

Batching also ushered in a streamlined affiliate-review process. HT worked

closely with J&J to forecast workflow as much as possible. Knowing that
10,000 words were scheduled for receipt at HT on a given date, HT could
schedule resources, and inform J&Js affiliate reviewers when they could
expect that batch for approval.
At the same time, HT implemented a standardized review process. Reviewers
would receive packages of approximately equal volume and format, and
thus grew comfortable with the look and feel of the text. They were
given standard instructions on how to mark up translations, reinforcing that
they were part of a process. This further reduced affiliate-review politics.
There is no shortcut for creating credibility in a process, to say nothing of
creating belief in translation quality among in-country reviewers. Review-

Robert C. Sprung 177

ers in a country who had become used to seeing mediocre translations

could hardly be expected to suddenly see the light that translations were
now on track. Credibility had to be built up over time, but proceduralizing
the review was an important factor in accelerating affiliate buy-in. The procedures also helped set common expectations, without which complaints
about quality are far more prevalent.
One improvement was instituting standard means of recording affiliates
comments, including a form for their global preferences in style and terminology (e.g., the French affiliates desire for using a certain term for extra
large when referring to an orthopedic device, or the wording for a standard warning statement). One should hardly accept at face value statements
such as the Czech reviewer said the translation was bad. Such forms
helped localize the source of the commentsthey might be politically
motivated (was the translation that used to be done in their country now
being outsourced to an upstart US firm?), or caused by procedural flaws
rather than by poor translators. (One phenomenon that emerged over the
years was a change in reviewer personnela Korean reviewer might have preferred a given term, and his replacement three years later might have been
upset at the existing mistranslation requested by the predecessor.)
Much of the solution rests in setting the right expectations: we began by
advising J&J that they should expect reviewers to be skeptical at first, that we
would have to win their approval. The decision to centralize, batch, and
proceduralize affiliate review yielded impressive results:

Previously, it had taken up to eight months to complete a standard,

2,000-word insert.

Now, it took a predictable six weeks for batches of up to 10,000 words,

which included multiple inserts and labels.

Ch al le nge 2 : C o nt rol l in g Tra ns l at i o n Q ual i ty

With a basic process in place for creating foreign-language text and artwork
in a predictable timeframe, J&J moved to a more subtle question: How can
US headquarters evaluate the quality of foreign-language labeling? To put it
in terms to which we can all relate: How do I know that my Japanese label
isnt saying something wrong or offensive, when I dont speak Japanese?
Because so much is riding on the accuracy of labeling, regulatory experts
require robust systems, typically ISO- or GMP-compliant, for generating
the text associated with a product. In their view, there is a clear right or
wrong in labeling text. On the other side are individual linguists or translation companies, which traditionally see more grey areas in a text. A good

178 Mission-Critical: Translating for Regulated Industries

translator is a good writer, and writing involves substantial creativity and

right-brain activity. How does one bridge this gap, in an era where ISO or
similar compliance is deemed a prerequisite?
Further complicating the matter is the very nature of ISO, which was
designed for a manufacturing environment. For example, ISO 9000 Section
4.11 covers Control of inspection, measuring, and test equipment.
Clearly this section and others are difficult to apply to a language solution.
In the end, J&J needed a language process that dovetailed with its own labeling procedures, and that could be understood by reasonably intelligent, but
non translation-savvy, regulatory auditors. The language vendor would ideally demonstrate a structured framework, consisting of the following:

a standardized, auditable process, integrated with the clients workflow;

resources that are properly qualified;

close integration with Regulatory Affairs;

project components that are fully traceable, in case of an audit;

a feedback mechanism for improving the process.

Sol uti on: A Sta nda rd ized Proce ss

The solution was to treat quality as a process, with standard procedures and
quantitative measures. The alternative was untenable over the long term
namely, an anecdotal approach to quality, where the parties get comfortable with translation style and accuracy. A translation vendor might sail
along without significant customer complaints for a given period, but ultimately the lack of a commonly understood and agreed standard with quantitative measures will come back to haunt both parties.
We agreed with J&J that the ISO quality framework would be most suitable
for their language solution. Gearing up for ISO compliance entailed a substantial investment of time and money on both sides, but this longer-term
commitment proved modest in comparison with the return.
As part of its own ISO or GMP certification, a company typically audits its
language vendors (under ISO 4.6, Purchasing, or comparable section in
another standard). Below are select ISO 9000 sections with some notes on
the criteria we applied, and which any firm may use in evaluating and
auditing a translation process. Although ISO was designed for manufacturers and not language-service companies, it can help control the foreign language on many firms labeling and documentation.

Robert C. Sprung 179

Section 4.3, Contract Review

Many problems in translation go back to poor communication and specification of scope of work. Typical problems: lack of specificity of deliverables
and lack of communication on technical details of delivery (e.g., a client
might have asked for Japanese text files, but did not understand that a Japanese operating system is needed to read the files).
Section 4.6, Purchasing

Under this section, clients audit their suppliers, including language vendors.
The same section covers how those language vendors qualify their own linguistic resources. Given the worldwide web of translators, it is critical to
document how a vendor qualifies any subcontractors who are not in-house.
Acid-test: a client should pull individual vendor files. Do they document
the proper level of education and experience to understand difficult labeling
text and translate it accurately?
Section 4.8, Production Identification and Traceability

The paper or electronic trail of a project is specified and stored under this
section. If there is ever a problem with foreign-language labeling, there
must be a way of tracing where the error occurred, when, and by whom.
Acid-test: a client should select an old labeling project and ask to see the
record of each step, from initial translation to final layout.
Section 4.9, Process Control

This is the trickiest section to implementthe heart of the quality system.

The ISO standard provides little explicit guidance, as the actual language of
the section is very brief and general: the production, installation, and servicing of processes which directly affect quality. In auditing a companys
process control, perform this acid test: does the supplier have a clearly documented process that is easy for the layman to understand? Does the process
include cross-checks? Who controls the process and what are these peoples
qualifications? In an audit, a client should talk with a translation project
manager to validate written procedures.
Section 4.13, Control of Nonconforming Product

It is easy to imagine an assembly line where a widget does not meet a certain threshold value for tensile strength. But can a translation be pulled from
a figurative assembly line for failing to meet critical values? Some companies
have devised ingenious methods for an analogous approach.2 If the editors
find problems in the translation, how is their feedback handled? Can substandard work be returned for rework? How is this documented? Note that

180 Mission-Critical: Translating for Regulated Industries

nonconformances include not only errors in translation, but also any nonadherence to a stated procedure that affects quality.
Section 4.18, Training

This is one of the most critical sections. It is not sufficient for a translator,
for example, to be well schooled and experienced. She or he must also be
trained for the task at hand. Acid test: a client should ask to see documented
requirements for an individual task (e.g., translation, editing, layout), as well
as the training records that show that an individual is qualified. Under a separate section, a language supplier must have internal auditing of its own systems. Are internal auditors qualified? (Pull their training records.)
Section 4.20, Statistical Techniques

At first blush, this section seems to have little to do with translation. However, a client and vendor should consider whether meaningful metrics
might not be devised, particularly if they have a large volume of foreignlanguage copy. What is the error rate per 10,000 words? How is throughput
measured? Together with J&J, we developed standard quality metrics,
included error rates per given volume of translation, with a classification of
errors (nonconformances) as major or minor. By gathering these data
over time, we could track trends in quality.
Ch al lenge 3: Da taba se La bel ing

With a basic process in place, JJPI next had technical problems to solve.
Specifically, how could its label design, storage, and printing systems be
optimized for a multilingual environment?
The company, which prints thousands of medical labels within its facilities
throughout the US and the world, now had to print foreign-language text
on those labels. How would quality be controlled on those labels? How
would existing printers output Japaneseand how would an operator
know whether the label had printed properly, or with incorrect characters?
J&Js wish list included:

support for foreign character sets (for languages currently needed, like
Finnish and Japanese, and for languages with complex character sets
that might be added, such as Czech or Korean);

2. Consider the Quality Assurance model proposed by the LISA organization, outlined in a
separate case study in this volume.

Robert C. Sprung 181

a relational-database structure (a flat-file structure, such as a single Excel

spreadsheet, creates inefficiencies through redundant data, and increases
the potential for error by having many translations where one will do);

reduction of overlabeling (the need to overprint or apply

supplementary labels in certain target markets).

Sol uti on: A Cu stom Ap pl ica t ion

The solution was a custom application that HT designed and programmed.

Its features may interest many in labeling or regulatory fields; more generally, it illustrates that custom technology solutions are often needed to solve
complex language problems.
J&J wished to retain its Zebra-brand printers, with which it had printed
labels for years. A software solution would save hundreds of thousands of
dollars in new hardware. In the end, HT provided an on-demand printing
system with the following features:

The system supported 12 languages, including Japanese and other

double-byte languages.

The system was multiuser: many could use it simultaneously, whether

entering labeling text, reviewing translations, or printing labels.

The system allowed remote access: label operators around the world
could log in and print labels.

The database structure was normalized and relational, using an off-theshelf application. This eliminated redundancy (for example, if the term
Store above 20 C appeared on 10,000 product codes, it was stored
once in the databasewhich would not be the case with an old-line,
flat-file structure). This made it much easier to control quality on
translations (put all your eggs in one basketand watch that basket),
while dramatically cutting translation cost.

Foreign languages are preserved as live text, not static artwork. Many
systems store double-byte text as art files (EPS, BMP, or PDF)
embedded into the database structure. This freezes the artwork,
preventing manipulation by a labeling operator. On the downside, any
change to a term in the database would require new artwork to be
created for every piece of Japanese text that might appear on a label.
For a company with 15,000 discrete labeling elements, this might cost a
small fortune in artwork expenses alone.

182 Mission-Critical: Translating for Regulated Industries

Label text goes through the same quality process as insert text, with a
complete paper trail and J&J-internal approvals. We found that many
companies have one team dealing with packaging text and another
with instructional inserts. We have already mentioned the potential
danger of not integrating translation of the twoinconsistencies of
such basic information as the products name can arise. Text that
appears on packaging is the most critical in terms of quality and recall
potential. Its translation, though not extensive in volume of words,
requires the same rigor as translation of a 2,000-word insert.

The system produced output in all languages (including Japanese and

Greek) to a Zebra labeling printer, including barcodes.

C A S E

S T U D Y

E T H I C O N

2 :

E N D O - S U R G E R Y

( E E S )

Ba s e li n e S it u at i o n ( 1 9 94 )

Ethicon Endo-Surgery, based in Cincinnati, Ohio, is a spin-off of Ethicon,

Inc., the worlds leading supplier of sutures. EES specializes in state-of-theart endoscopic and laparoscopic instruments.
The differences between the EES and JJPI corporate structures and growth
strategies had profound implications on translation requirements. JJPI,
while rolling out some new products, had a product line of million-sellers
that had sold well for decades. EESs growth was fueled primarily by newproduct development. This meant even greater pressure on translations
every week saved could have a major impact on product sales. Hence, translation was viewed from the outset as having a major strategic role for the
company.
Another complicating factor was EESs aggressive acquisition strategy.
When a company acquires another, it faces these basic language issues:

Does the acquired company require translation? If little has been

translated, a large volume of new work is possible, since management
often needs to rapidly bring both companies product lines into
conformity, which includes translation.

Does the acquired company have existing standards for translation? In a

worst-case scenario, the acquired companys translations are poor,
produced without the acquirers rigorous processes, and thus requiring
substantial rework. Even if translation had been done well,
inconsistencies in terminology, style, and graphics are inevitable,

Robert C. Sprung 183

necessitating at least a careful review of any translated materials from the

acquired company.
Ch al le nge 1 : C ommuni cati ons

EES benefited from the translation processes already developed for J&J. Still,
they desired to explore whether time could be cut from the process in the
age of such technological breakthroughs as the Internet.
One must keep in mind the nature of regulated industries. The medical
industry treads very cautiously when it comes to replacing traditional practices such as a paper trail and handwritten signatures; the regulatory apparatus will not adopt such innovations as electronic signatures or a virtual paper
trail until they are rigorously tested and foolproof.
Nonetheless, each existing means of communication presented an obstacle
or bottleneck:

Fax was deemed unreliable, messy, and time-consuming.

FedEx or similar overnight service was expensive (a company could

spend hundreds of dollars a month shipping text for review to 10
affiliate offices). Further, at least two days were lost in each direction.
Although FedEx was reliable, if there were six shipments in a month to
10 affiliate offices, at least one or two packages would experience some
sort of delivery problem, yielding painful delays in the final translation.

Email presented security and conversion problems, particularly in the

early days. File formats remain a challenge when people use programs
on different platforms (including Mac-to-PC issues, font problems, and
PDF format issues).

Sol uti on: A Corp orate Tra nsl ati on In tran et

EES, working with HT, developed a state-of-the-art technology solution.

The Global Translation Process System, or GTPS, is a corporate intranet,
housed at EES headquarters, that serves as a central clearing house for the
entire translation process.
GTPS has a central server that allows EES, its affiliate reviewers, and HT to
store, upload, and download project files. It grants varying access levels to
the parties involved. The system tracks and controls all project componentsindispensable in a mission-critical activity where all parties, from
translators to reviewers, must be held accountable for performing on schedule. Here are some key system features:3

184 Mission-Critical: Translating for Regulated Industries

State-of-the-art global approval. EES affiliates log directly into the system,
and work in a common file format. This solved several problems. First,
cumbersome and expensive overnight deliveries were a thing of the
past. (By some estimates, these savings alone rationalized the entire
programming effort.) Second, the standard file formats and conventions
for modifying and marking up translations ensured minimal problems
with conversion, diacritical marks, etc. EES technical staff worked with
affiliates to ensure that they had common hardware and software. Third,
the polished GTPS interface and process solidified relationships with
the affiliates and heightened their confidence in the translation process.

Real-time communications between headquarters/affiliates/translation vendor.

Files were transmitted instantaneously between all parties. There were
suddenly few, if any, transmission errors, while previously any FedEx
glitch added significant time tracking the problem, making new copies,
alerting parties, and adjusting schedules.

Published schedule and approval status. A major problem had been

adherence to schedules. By publishing schedules and tracking results,
EES reined in the process and held people accountable. In the past, if an
affiliate had said review time was insufficient, the company had no
recourse. Now, EES was collecting performance data, with which the
company could truly manage and control the process.

Official glossary continually updated and available. A key to good translation

in regulated industries, as we have seen, is consistency. Terminology
must be kept in a continually updated glossary. By posting this live
document on an intranet, all parties can access the same terminology.

Ch al le nge 2 : C o nfl ic t i ng P r i o r it i e s

The single greatest challenge in translating for highly regulated industries is

arguably the following: On the one hand are forces demanding economy
and efficiency (do it faster, cheaper); on the other are equally powerful
forces for quality (embodied by Regulatory Affairs). We stress powerful
when discussing Regulatory Affairs. These are the people whowhether
inside the client company or from an outside agency like the FDA or
SECcan literally shut down an operation that they feel is not in
compliance.

3. A more thorough description can be found in a story about the GTPS and Ethicon EndoSurgery published in Pharmaceutical and Medical Packaging News, ISSUE January 1997.

Robert C. Sprung 185

The basic question is: How do we reduce translation cost and turnaround
time while satisfying Regulatory Affairs?
Solution: Process Integration

The solution rested in applying a process orientation, as discussed under

Standardized Process above, but going one step furtheractually ingraining translation in the organization so that all parties, including Regulatory,
have a say in work processes, and understand what will be happening (as
opposed to a reactive mode, where people are suddenly confronted by
translation outputs).
In the proactive, strategic model that EES and HT jointly developed, these
were some of the benefits:

Work forecasts became more common, so that all parties became aware
of how translation worked, and how to budget and schedule for it.

Translation became more integrated into the production cycle, with

technical writers who produced the original English as well as
engineers who designed the product starting to think more
internationally, thus simplifying translation and rendering the final
product more user-friendly around the world.

There was greater information-sharing among all parties. Rather than

fostering an antagonistic relationship, the processes and systems
reinforced cooperation. Audits were collaborative efforts that helped
educate the parties that were new to translation.

Regulatory Affairs became more comfortable with some of the

cutting-edge solutions, and also had significant input in them. For
example, parts of the review process became essentially paperless,
while preserving electronically a rigorous virtual paper trail. This
safeguarded the critical traceability required, for example, under
ISO 4.8.

Regulatory Affairs function became proceduralized and standardized.

Rather than acting as an arms-length third party with little knowledge
of the process (common at many companies), Regulatory was involved
early, minimizing 11th-hour surprises and allowing translation to
proceed more smoothly.

All other roles were clearly defined, including that of the translation
vendor and the affiliates. Explicit roles meant fewer surprises, with clear
goals and accountability.

186 Mission-Critical: Translating for Regulated Industries

Su mm ar y

The results of the above process and technology improvements have been
substantial. Qualitatively, they included:

reduced time-to-market;

reduced costs;

improved quality;

improved communications.

But these advantages were not left in the realm of the anecdotal and qualitative. Most regulated industries share a drive to develop metrics and quantify
results for key systems, and this was also the case with the language solutions
implemented here. The following are some manifestations of the strategic,
rather than tactical, view of language that J&J embraced:

Key procedures for translation were documented, using an ISO- and

GMP-compliant quality system.

Quality metrics were applied to the language solution, such as tracking

nonconformances, with an appropriate mechanism for taking
corrective action.

Performance data were recorded and analyzed, including standard

turnaround, throughput, and efficiency of all parties.

Given tremendous time-to-market pressurewhich will only get more

acute with the arrival of the Internetthese solutions represent a sane
approach for the company going global that sees overcoming language barriers as a strategic advantage.

13

L A N G U A G E A U T O M A T I O N
C A S E S T U DY

Section V: Language Automation

13

Machine Translation and

Controlled Authoring at Caterpillar
Rose Lockwood
Equipe Consortium Ltd.

S U M M A R Y

Caterpillar is a pioneer in the implementation of machine translation based

on controlled-English documentation. An already huge publishing volume
is expanding as the company globalizes into new markets. To handle
increased volumes, more languages, and the need for electronic distribution, Caterpillar has completely reengineered its documentation and
publishing processes, implementing an innovative system combining SGML
databases with automatic translation.
C O M P A N Y

B A C K G R O U N D

Caterpillar, based in Peoria, Illinois, USA, is a growth-oriented, high-tech

global company, and the worlds leading manufacturer of construction and
mining equipment, natural gas engines, and industrial gas turbines. It is a
leading global supplier of diesel engines and also offers varied financing
options through its Financial Products Division. The company had 1997
revenues of US$18.9 billion, with US$1.7 billion profits.
Throughout the 1990s, Caterpillar has been engaged in a corporate reengineering program aimed to create a decentralized, function-oriented
organization. Part of that strategy is to operate with a global perspective,
and to target growth markets worldwide, particularly in emerging markets
with high demand for construction equipment. About half of Caterpillars
sales are outside the US (51 percent in 1997).

1. This case study is based on material previously published in: Lockwood, et al., Globalisation:
Creating New Markets with Translation Technology (Ovum, 1995); Kamprath, et al., Controlled
Language for Multilingual Document Production: the Experience of Caterpillar Technical
English, Proceedings of the CLAW 98 Conference; Nyberg, et al., The KANT Translation
SystemFrom R&D to Large-Scale Deployment, LISA Newsletter, March 1998. The latter two articles can be viewed at [Link]/Research/Kant (research articles number 20 and 21). The opinions expressed in this case study are those of the author, and not of
Caterpillar or CMT/CMU.

188 Machine Translation and Controlled Authoring at Caterpillar

Caterpillar has a significant manufacturing presence outside the US, for better access to its key markets. Manufacturing sites include:

Australia
Belgium
Brazil
Canada
China
France

Germany
Hungary
India
Indonesia
Italy
Japan

Mexico
Northern Ireland
Poland
Russia
Sweden
United Kingdom

Caterpillar pursues its globalization strategy in partnership with companies

located in emerging markets. For example, joint-venture agreements have
been reached with AMO-Zil and Kirovsky Zavod in Russia, and Shanghai
Diesel and Zuzhou Construction Machine Group in China. With nearly
half the worlds mineral wealth, these two countries are important markets
for Caterpillars products. Countries where Caterpillar operates have 74
percent of the worlds land area and 81 percent of its population; the company has a strong position in infrastructure development and electrification
of emerging markets. Despite difficulties in Asia in 1998, the company has
continued to show strong sales.
Caterpillar has established a strong product-support system to service its
dealership network worldwide. Extending to nearly 200 countries, the
company considers the network a key competitive edge, and meeting the
language needs of that network has been one of the drivers of the translation-automation program. Most of the companys 197 dealerships are
locally owned independent businesses, and many dealers have been with
Caterpillar for 50 years. This network operates in over 35 languages; providing product information in local languages is a strategic issue for
Caterpillar as it pursues its globalization strategy.
C A T E R P I L L A R S

D O C U M E N T A T I O N

N E E D S

Caterpillar produces a large volume and variety of technical publications,

including documentation for over 350 current products and older products
still in use. Documentation must be available for the life of a product, which
tends to be long. Each Caterpillar product incorporates several complex
subsystems, such as an engine, hydraulic system, drive system, implements,
and electrical systems, for which a variety of technical documents must be
produced. They include operation and maintenance manuals, testing and
adjusting instructions, disassembly and assembly manuals, and specifications,
as well as bulletins regarding service procedures and new service offerings.

Rose Lockwood 189

This information is provided to all dealerships. Moreover, the volume is

always increasing. Many dealers have found it difficult to locate the information they need, when they need itwhether at the parts counter, in the
service bays, or in the field where equipment is often serviced. To solve this
problem, Caterpillar launched a project called the Service Information System (SIS), an electronic delivery system supplying documentation to dealers
on CD-Rom.
In non-English-speaking markets, dealers require as much documentation
as possible in translation. Therefore, in developing SIS, Caterpillar planned
to improve its translation program and increase the amount of information
supplied in local languages. Documentation is routinely translated into
Spanish, French, German, Italian, Portuguese, Danish, Dutch, Finnish,
Swedish, Norwegian, Greek, and Russian. Some material is also translated
into Hungarian, Thai, Arabic, Chinese, Japanese, Icelandic, Turkish,
Bahasa, Vietnamese, and Korean. Caterpillar now runs a substantial publishing operation, generating over 800 pages of new material a day and
shipping over 110 tons of printed materials to dealers every month (since
some material is still required on paper). The company maintains the largest
CD-Rom library in the world, and generates over 50 CD-Rom masters
every month.
Product-development cycles are getting shorter at Caterpillar, putting additional strain on the documentation system. Five years ago, the entire
product development cycle was 18 months to two years; it is now six
months or less.
Documentation requirements at Caterpillar can be summarized as:
large volumes with frequent updates;
electronic delivery and print formats;
multiple-language versions;
shorter documentation cycle times.

190 Machine Translation and Controlled Authoring at Caterpillar

T H E

C A T E R P I L L A R

S O L U T I O N

Caterpillar has made a long-term investment in innovative technologies and

processes to solve language problems. The current solution has been in
development for a number of years, and is based on the use of controlledEnglish authoring with Caterpillar Technical English, an SGML database
for file and document management, and automatic translation. The Caterpillar system has been granted a US patent, and international patents are
pending.2
Ca terp il l ar Te chn ica l Engl i sh and th e La nguage Env ironment

The company began to face the problems of a multilingual customer base in

the early 1980s. This problem was originally addressed through the creation
of Caterpillar Fundamental English (CFE)a set of around 1,000 English
words which authors attempted to use in all documentation. The company
developed a CFE training program for overseas dealers and staff to eliminate
the need to localize documentation.
Over time, it became clear that CFE was not a total solution. Authors were
forced to use elaborate and lengthy descriptions of concepts that were not
in CFE, and English users sometimes found the material unreadable. As
Caterpillars product base became more complex, adding high-tech features
such as high-pressure hydraulics and electronics, the limited vocabulary of
CFE became increasingly inadequate for technical publishing. High staff
turnover and low levels of English literacy in the field-service staff made the
increasing complexity even more difficult to support with CFE training
alone. As Caterpillar moved into more markets, some with languages using
non-Roman alphabets, and a growing number demanding product localization, it became clear that more translation was required.3
Caterpillar has reestablished controlled English as a technique which supports both document management and the translation needs of the
publications department. In researching options for automatic translation to
meet the ever-growing volume of translated material, it became clear that
having a constrained set of inputs would greatly improve the possibility of
achieving high-quality machine-based translation. In the early 1990s the
company launched a project with The Carnegie Group, Inc. (CGI in Pitts2. US Patent 5,677,835Integrated Authoring and Translation System, Carbonell et al.,
granted October 1997covering a wide range of functions of the system.
3. Use of CFE was discontinued by Caterpillar in 1982. The technology was subsequently
licensed by Smart Communications, who have developed it further for their clients; the
Communication Studies Unit of the University of Wales Institute of Science and Technology
have also used CFE in R&D projects. See Nyberg, et al.

Rose Lockwood 191

burgh, Pennsylvania) to develop controlled-English documentation and

automatic translation into a subset of their 15 target languages. CGI served
as a systems integrator for linguistic tools developed at Carnegie Mellon
University (CMU).
Rather than attempting to limit text to an extremely small number of terms
(as had been the case with CFE), the company started from scratch to
reduce all terms and phrases to a standard, if much larger, set. They selected
900 Caterpillar documents representative of all types of technical publications and all existing product types. Using linguistic analysis tools developed
by the Center for Machine Translation (CMT4) at CMU, CGI parsed this
large volume of material and reduced the text to some one million unique
words and phrases. Caterpillar staff then pared this set down, eliminating
redundant terms (where different authors had used different terms for the
same concept or object), and reduced the set to approximately 70,000.
This set of terms, along with a grammar that defines acceptable usage, is
Caterpillar Technical English (CTE), and forms a fundamental part of Caterpillars multilingual solution. CMT and Caterpillar staff worked together
to devise guidelines for using CTE in authoring, and wrote usage examples
for each CTE term. CTE consists of these 70,000 terms and a large subset
of the structures of standard English in the CTE grammar.
Over 100 technical authors access CTE from a central server, using a CGIdeveloped Language Environment (LE) interface. The system provides a
full linguistic sentence-by-sentence analysis of each document written by
the authors. The LE user-interface provides definitions of terms and usage
examples to authors; it controls the use of terms and provides interactive
disambiguation to assure consistency and accuracy in English source text,
and more accurate automated translation. The LE software interacts with
the machine-translation engine as a co-process, while managing interactions with the user and the SGML text-editing environment, ArborText.
Initial training for CTE authors began in 1994, with bimonthly brownbag seminars throughout the year in advance of initial CTE implementation. Formal author training began in 1995. The entire staff of authors was
brought into the system gradually, while document types have been added
incrementally. Standard training for authors has included:

two weeks training to use ArborText (and its Unix environment), the
file-management system, and graphics tools;

4. The CMT was expanded in 1996 to become the Language Technology Institute, LTI (see
[Link]).

192 Machine Translation and Controlled Authoring at Caterpillar

three weeks of mentoring by a CTE expert.

SGM L D atab ase for Fi le Ma nagem en t

While the CTE solution was being developed, the Caterpillar publications
department began to develop a text- and document-management system
that would support the new electronic delivery systems needed for SIS. The
old document formats, based on a proprietary generic coding system developed many years prior for input to an ATEX typesetting system, were not
easily converted to formats appropriate for CD-Rom and electronic publishing. Caterpillar decided to develop the new system as an SGML database
to generate both print documents and elements for CD-Rom and online
delivery.
SGM L Descri pti ve vs. P resen tati onal Codi ng

Proprietary text-editing formats (such as those used by word processors or

desktop-publishing systems) address the layout and appearance of a document on a page. This limits their usefulness as coding schemes for
documents which will be presented in other formats (help screens, for
example)and especially for documents that may be presented in a number
of different formats, such as Caterpillar requires. An alternate approach to
document coding is descriptive, and much more akin to database coding. It
describes the structural elements of a document such as headings, body text,
and sections. The most commonly used descriptive coding system is Standard Generalized Markup Language (SGML).
SGML was published as ISO 8879 in 1986, and has since been adopted by a
number of vendors (including a longer-term commitment from Microsoft)
and, most significantly, was incorporated into the CALS (computer-assisted
logistical support) standard required by the US Department of Defense for
all documentation in systems supplied to them. This requirement has been
extended to all US-government tendering documents, ensuring that SGML
will be an important standard.
SGML is a tagging methodology whereby the structure of any text (including graphics and other nonlinguistic elements) can be maintained
independently of the devices and software used to create the document.
Parallel and nested information structures can be maintained, and can be
format- and content-related, so that SGML actually extends the concept of
markup into new areas.
Strictly speaking, markup is everything in a document that is not contentrelated, and the term comes from marking up typed text with instructions
for the typesetter or compositor about aspects such as how to fit the text on
a page, what typeface to use, etc. SGML takes this generic concept a step

Rose Lockwood 193

further and specifies a method for setting up document-hierarchy models

where every element fits into a logical, predictable structure. It includes a
syntax for defining document data elements and the overall framework.
SGML is not a set of tags or document typesit is a method for designing
tags and rules for any type of document. In other words, it is a programming language.
The heart of an SGML application is a document type definition (DTD) that
defines the structure of a document in the same way that a database design
describes the structure of a database. Elements defined by a DTD might
include chapters, chapter headings, sections, and topics. It also defines
rulese.g., A chapter heading tag must be the first element after a chapter
tag. The DTD accompanies the document wherever it goes in a publishing system.
The DTD also provides a framework for the content of a document. Tagging identifies the role of content within a document and tags mark the
beginning and end of each content type. Creating an SGML document thus
involves adding text and inserting tags around it. The DTD is written in a
rigorously defined language which is handled by an SGML parser. The
parser is used to develop DTDs and verify that documents conform to the
standard.
The elegance of the SGML solution is that it does not restrict document
definitions to any particular type of output, and this is one reason it is seen
as the standard of choice for electronic publishing. Output styles are defined
in formatting output specification instances (FOSIs), which are a special kind of
DTD. A FOSI is like a large style sheet which specifies the formatting characteristics for each tag in each of its contexts. Alternatively, one may use
proprietary editing systems for coding the descriptive structure provided by
SGML.
SGML turns unstructured informationwhich constitutes 90 percent of
information stored electronicallyinto structured data. In effect, document resources can be handled like data because they are in a database
environment. This has highly significant implications for publishing and
information provision, especially in multilingual environments.
In the world of online or electronic publishing, variants of SGML have
evolvedstarting with HTML (Hyperlinked Text Markup Language),
which is a standard DTD for use with hyperlinked Web pages, and is used
on most Web pages now published. This standard will, however, give way
to XML, a more sophisticated variant capable of handling the complex
multimedia documents published on the Web.

194 Machine Translation and Controlled Authoring at Caterpillar

Ca terp il l ars SGML Solution I nfor ma ti o n E l em e nt s

To support an SGML database, Caterpillar had to introduce a new authoring platform. The new system uses an SGML editor developed by
ArborText to create the DTDs and FOSIs needed for the new file-management environment. SGML editors facilitate definition of DTDs and FOSIs,
and creation of SGML-compliant documents. These editors provide a
friendly editing environment with visual cues about the structure of the
document, which removes the burden of knowing SGML programming.
The most revolutionary innovation in Caterpillars document-management
approach may well be the adoption of information elements to replace
documents in the publishing system. Instead of writing books or manuals,
authors now write information elements (IEs) that describe a specific concept or procedure. IEs are usually two or three pages long and are reused
whenever possible. An author only writes what is specific to a particular
information element. These elements are combined with other IEs and
with shared objects (which can be graphics, tables, etc.) to build books or
CD files as needed. This write once, use many approach produces substantial time and cost savings.
One requirement for CTE development was that it should support modular
creation of documents as part of these reusable information elements, to
improve consistency in English documentation as well as create a common
look-and-feel for Caterpillars technical publications. Since documents are
created by combining different IEs, the CTE system needs to make IEs
consistent, despite different levels of expertise and experience among
authoring staff. It needs to incorporate SGML tagging for both paper publishing and electronic delivery of multilingual texts, and its grammar needs
to handle complex and variable types of markup within information elements, including grammatical markup for target-language output in
translation.
These IEs are combined into one of the dozen different document types
identified for Caterpillar technical publications. Each document type has a
DTD that defines the internal document structure in English and 15 other
languages, and includes general style guidelines.
A File Management System (FMS) is a system that controls the creation,
update, and translation of IEs. It is an integrated set of document-management tools (incorporating some 60 different software elements), managed
by a custom-built workflow-management tool. Using workflow management, IEs are passed through tasks assigned to various authoring and
translation staff.

Rose Lockwood 195

A ut o ma t ic Ma c hi ne Tra ns l at i o n (AM T ) Sy s t em

With expanding volumes of translation at Caterpillar, a machine-based

translation system was an integral part of the solution. However, a review of
available systems revealed none that would provide adequate quality without the need for extensive post-editing. Caterpillars goal is to do as much
pre-editing as possible at the authoring stage of the document cycle, to
reduce the time and costs of translation. They decided to work with the
CMT at Carnegie Mellon University to develop a custom solution to
exploit the benefits of controlled authoring. The system they developed is
Caterpillars AMT system, based on CMTs Kant technology.
Over vi ew o f K nowl ed ge - Ba s ed A c c u r at e N at ur al - Lan guage
Tr ans l a ti o n ( Ka nt )

CMT initiated the Kant project in 1989 for the research and development
of large-scale, practical translation systems for technical documentation.
The Kant translation engine uses a controlled vocabulary and grammar for
each source language, and explicit semantic models for each technical
domain, to achieve very high accuracy in translation. Designed for multilingual document production, Kant has been applied to electric power-utility
management in addition to heavy-equipment technical documentation at
Caterpillar.5
Kant is CMUs MT system for automatic translation in technical domains.
Kant is dubbed knowledge-based MT because it makes use of lexicons,
grammars, and mapping rules for each language. Users can define terminology for specific domains, where the meaning of terms is limited by
relevance. This approach tends to produce more accurate translations of
technical material than general-purpose MT systems and the output typically requires modest post-editing.
Kant supports use of controlled input language with explicit conformancechecking during text authoring, applying the same grammar used for translation. It also supports the use of SGML tagging within sentencesSGML
is treated as an integral part of text to be translated.
Caterpillars translation system was based on the original Kant architecture.
However, the Kant development team at CMT is redesigning the software;
the new version is called Kantoo (Kant object-oriented). Caterpillar is
gradually replacing the initial modules of their system with the reengineered modules of Kantoo. The main purpose is to improve
5. CMT and CMU are forming a new commercial entity, Alantra Systems, Inc., which will
be the primary vehicle for marketing Kant technology.

196 Machine Translation and Controlled Authoring at Caterpillar

implementation and deployment of new Kant applications, rather than to

change the language engines per se. Kantoo will include a Lexicon Maintenance Tool (LMT) to handle source-language terminology (i.e., CTE
terminology at Caterpillar), and a Language Translation Database (LTD) for
target-language terminology. Both tools are built on a PC-based Oracle
database and forms applications, and are designed for rapid development
and maintenance of linguistic resources.
The Kantoo Analyzer and Generator are reimplementations of the Kant
toolsthe Analyzer used for grammar checking and source-text analysis
during translation, and the Generator a reimplementation of the target-language translation engine. The Knowledge Maintenance Tool (KMT) is a
graphical user-interface
for real-time browsing, editing, and updating of
lexica, grammars, domain models, and mapping rulesi.e., all knowledge
sources used during analysis of source-language text and generation of target-language text. At present writing, development milestones are complete
for all five modules of the Kantoo system and are now being integrated into
the Caterpillar writing and translation environment.
The reengineering of Kantoo is driven principally by the need for better
software-engineering procedures, such as reducing the cost and effort
involved in maintenance through better database management. All elements
of the system are being ported from LISP to C++, to make the system
usable on more platforms, including standard desktop environments such as
Microsoft Windows, more robust in operational environments, and easier to
maintain.
I nt eg r at i ng A ut ho r i ng a nd Tran s la t ion To o ls

The AMT system contains a database (or knowledge base) of CTE vocabulary and phrases, acceptable syntactic constructions, and relationships
between concepts which constitute the Caterpillar domain knowledge.
One of the tasks of the original CGI contract was to develop the interface
tool which would give authors access to the linguistic processing modules
of CMTs system, including this domain knowledge. Thus the same grammar system used to analyze text for translation is also used at the authoring
stage to check whether new text conforms to CTE rules (for terminology
and grammar). Where material does not conform, the editor suggests alternative vocabulary or offers examples of approved usage.
CTE terminology, which is also the dictionary resource for AMT, is too
complex to control completely. Common terms with ambiguous meanings
are the most difficult to control. Caterpillar uses the example of the English
word valve, which can have at least nine different meanings depending on

Rose Lockwood 197

the type of valve and the role it plays in various systems. Similarly, the verb
charge can mean charging an electrical element such as a battery, or pressurizing a gas container such as an ether cylinder.
While CTE originally set out to define every term or phrase with a single
meaning (to assure accurate translation), this did not prove possible. Around
2,500 terms in CTE are ambiguous and may cause translation errors in
AMT. To overcome this problem, Caterpillar employs SGML codes to disambiguate these terms in the authoring process. When such a term is
used, the CTE software prompts the author to indicate the intended meaning, and this information is tagged in the text. A similar technique is used to
handle structural ambiguitiessuch as referents for prepositional phrases. At
the time of translation, the AMT system selects the correct meaning based
on the SGML codes.
After implementation of CTE, Caterpillar discovered that this disambiguation option could have negative consequences, where alternative meanings
are irrelevant in a particular type of document. If an author is writing an IE
pertaining to an electrical element, he or she doesnt want to be presented
with the pressurizing definition for charge as an option in that document. A refinement was introduced, to attach subdomain codes to the IE;
now only meanings relevant to the subdomain are offered in the interactive
disambiguation session.
New IEs are filed in the File Management System, and the English version
is subsequently submitted to the AMT system for translation. Authoring
and translation are strictly separate. Translation itself is in two stages:

the English version is analyzed and translated into a languageindependent representation called an interlingua

the target language is generated from the interlingua

The interlingua concept is central to the Carnegie-Mellon approach to

machine translation. The advantage is that once the interlingua has been
created, a new target language can be added by building a language generator for that language. No further parsing or analysis of the source text is
required, as is the case with other approaches to MT (such as the so-called
transfer method).
Both analysis and generation functions of the AMT system access the Caterpillar domain knowledge base. Each module accesses a base of
information about whatever language is being handled; modeling targetlanguage matches of CTE vocabulary is part of this language information.

198 Machine Translation and Controlled Authoring at Caterpillar

Lan guages I mp le me nt e d

The first AMT module became operational for French in 1996. With
AMT, the translators principal role is to check and do a final edit on the
results of the automatic translation (post-editing). Prior to implementation, Caterpillar undertook an acceptance evaluation for French AMT,
looking at various usability, performance, and quality factors. The result
determined that improved post-editing productivity was the primary goal of
the system. By some measures, productivity improvement for translators
showed as much as a two-to-one gain. Caterpillar is cautious, however,
about making definitive claims regarding productivity, since the new system
is complex, and touches almost every aspect of the document creation,
maintenance, translation, and publishing cycle. The results of the evaluation
were sufficiently positive, however, to justify acceptance of AMT for
French.
The next planned language, Spanish, was accepted for production use in
1997. Acceptance testing for Spanish was done somewhat differently.
Rather than using a formal usability study (as with French), they used a specialist team, including a Spanish expert linguist and the CMT developers, to
do rapid feedback and refinement over several months, in what was essentially alpha testing. Evaluation of Spanish in production mode is ongoing,
and translators have reported that Spanish output requires little post-editing.
Caterpillar is currently implementing the German AMT system.
Ma in ten an ce an d Costs

Caterpillar has gradually begun maintaining elements of the system. As of

late 1998, they had taken on maintenance of source-text data (English terminology and grammar updates). They have developed a problem-reporting
process, enabling authors to request term and grammar updates. Ongoing
maintenance of CTE terminology controls new terms, removes redundant
terms, and screens those requested by authors. Usage examples are also
updated whenever CTE grammar is improved, and these examples must be
revalidated.
Development of CTE, the LE interface, and target-language translation
engines (for French and Spanish) all took place in parallel. The development phase for the system was relatively long, and the company committed
five full-time equivalent staff per year over a five-year period. Participants
from Caterpillar included linguists, pilot authors, trainers, mentors, and system developers. These resources were devoted principally to initial
development of terminology and grammar, evaluation and refinement, and
training. Ongoing operational costs include incremental updating of termi-

Rose Lockwood 199

nology for new products requiring new documentation, and refinement of

the grammar as new document types are introduced or existing ones
revised.
C H A L L E N G E S

Caterpillars new authoring and translation system has involved substantial

reengineering of internal processes, as well as development and implementation of new technology, and no change of this magnitude can be troublefree. Because various elements were developed in parallel, one challenge has
been integrating the modules.
The CTE domain is too complex for lexicographers to anticipate all the
ways authors will use terms. Since ambiguities cannot be defined in
advance, terms and grammar must be constantly extended. Post-implementation support will be substantial, and the company will need to learn to
shift skills from traditional translation roles to the more technical tasks of
maintaining and interacting with the system.
The Caterpillar solution represents a complete change in the working paradigm of the publications division, and there has been understandable fear
and worry on the part of authors because their jobs are now significantly
different. Nevertheless, most authors have found the new system easier to
use than they expected, and appreciate the ease with which they can pick
up IEs to assemble new publications. Despite the sophisticated features of
the CTE-checking tools, compliance by authors can be variable or difficult
to enforce. Authors may use words with the wrong meaning, or select the
wrong meaning during disambiguation. It is also possible to write syntactically correct sentences that are incomprehensible or inaccurate. Moreover,
there are situations in which full CTE conformance is suspendedsuch as
deadlines for product launches which preclude full coding of new terminologyand these can be used by authors to minimize the control of CTE on
their writing.
In the course of implementing IE authoring with CTE checking, it became
clear that sometimes an author would reuse an information element with
only slight revisions, but the whole IE would have to be reparsed by CTE
for compliance. A region marking feature has been added, to mark areas
of text (usually sentences) that are CTE-approved and unchanged. This,
combined with a batch-checking application for modified IEs, can quickly
return only those sentences that fail to pass CTE, or require
disambiguation.

200 Machine Translation and Controlled Authoring at Caterpillar

Updating the CTE dictionary is a continuing challenge as new products

emerge and little time remains between final engineering and shipment
date. Unknown terms can now be temporarily protected from CTE analysis, and an automated process for requesting and screening missing terms has
been introduced. This assures conformance of major parts of new texts,
without requiring immediate new-term authorization.
The effort required for terminology work (especially during development)
is usually underestimated. The ongoing maintenance load for CTE and target-language term development will naturally fluctuate, and be difficult to
plan. Establishing translation equivalents for all new CTE terms requires
ongoing vigilance to maintain consistency in the databasesa burden
which will increase as new languages are added to AMT. Ambiguity in the
English text can now be handled by coding during terminology development. As new AMT languages are implemented, all of which translate the
same CTE input, it will become necessary to adjust the relation between
English and target-language terminology.
Like many innovators in language engineering, Caterpillar finds it challenging to locate adequately skilled staff for some of the technical work,
especially in terminology coding. The task needs a combination of detailed
knowledge of the CTE domain, training in linguistics, and a thorough
knowledge of the target language (i.e., translator skills). Caterpillars solution requires highly skilled personnel at all levels, from authors and
translators to terminologists, lexicographers, and system maintainers. In the
medium term, they will probably continue to use outside services (such as
those of CMT) to complement in-house skills.
Translators are being asked to move from traditional roles, where they are
generally in control of the style and content of translation, to that of posteditor, and some are not enthusiastic about the change. The learning curve
for effective post-editing is not trivial, and it implies a level of retraining for
translator staffa particular challenge where outside translators are used.
The level of appropriate post-editing is still hard to define; editing to a standard of stylistic consistency is difficult to do efficiently (and not necessary
where generated text is understandable), and some translators are reluctant
to do minimal revision. Managers of translators will need new ways to measure translator effectiveness and make judgments about quality.
Finally, it is still a challenge to devise metrics for evaluation of AMT which
are not influenced by factors outside the translation process, such as author
compliance with CTE, the quality of the term database, knowledge of
SGML, and system training. In the long run, Caterpillar will need to

Rose Lockwood 201

address the process improvements of the entire publishing system, as all elements in the document cycle are brought within an integrated
environment.
Benefits and Str ategic Dire ctions

CTE has delivered benefits for both manual and automatic translation at
Caterpillar. While consistent source-text authoring has been implemented
principally to improve the output consistency of MT, it also improves the
consistency and cost-effectiveness of manual translation. Controlled-English
authoring also improves the quality and consistency of documentation for
English-language users.
Caterpillars solution has allowed the company to begin developing a full set
of terminology resources for use in all company documentationa benefit
which was not initially a major goal for the system, but which may prove to
be among the greatest it confers, as all elements of corporate publishing
begin to leverage readily available and reusable terminology.
The implementation of the IE architecture has created substantial reusability
in Caterpillars publications database. IEs are now being reused across product lines, which increases production efficiency and cuts cost. The system
has essentially delivered an author-memory system which eliminates redundant rewriting. Authors are increasingly skilled at reusing IEs for compiling
new publications.
Implementation of CTE, IEs, and AMT have heightened awareness of language-related issues at Caterpillar. Close checking of authored text raises
the profile within the organization of issues relating to documentation quality, such as the consistency that comes from conforming to authoring
guidelines and standardization of processes.
Key benefits from the system include the following:

technical publishing is shorter and more efficient;

new document types that Caterpillar previously could not handle can
be translated;

documentation will be reusable in different formats, for different

purposes, leveraging valuable corporate resources;

electronic distribution of documentation will be simple and integrated

with the mainstream processes;

technical writing will be more closely integrated with production, as

authors will be part of the new-product introduction team;

202 Machine Translation and Controlled Authoring at Caterpillar

products and translated documentation will ship simultaneously (in the

past, translated versions lagged by eight to 10 weeks).

Although the system has required substantial investment and several years of
development effort, overall costs for documentation will be lower, particularly considering the number of languages that will be handled. At the same
time, costs will be substantially redistributed towards tools that support controlled authoring and editing, and management of documentation elements.
Expansion to other languages, particularly Asian languages, is under investigation. As an interim solution for other languages, Caterpillar uses
translation memory, particularly for high-reuse documents such as
updates of manuals. This remains the long-term solution for languages not
automated through AMT.
The introduction of a multilingual document-management system has revolutionized publishing at Caterpillar. The company believes this system will
give it a significant competitive advantage as it continues to pursue new
markets worldwide.

14

L A N G U A G E A U T O M A T I O N
C A S E S T U DY

14

Combining Machine Translation

with Translation Memory at Baan
Carmen Andrs Lange
Baan Development B.V.
Winfield Scott Bennett
Logos Corporation
I N T R O D U C T I O N

The demand for translation is currently rising dramatically, given the boom
in international trade. Pressure on translators and translation companies to
produce high-quality results in shorter times is likewise increasing. A major
problem is the lack of qualified translators to meet the rising need. The
dream of applying computer technology to assist in the translation process
has been one of the earliest goals in the field of artificial intelligence. Long
turnaround times for translation coupled with demand for translation services that outstrip the supply of human capital have again thrust technology
into the forefront of the conversation.
What was once the realm of hype and fearwill machines replace human
translators?has emerged as a viable partner for the translation industry.
Translators now routinely use a variety of tools, ranging from basic word
processing and database-management systems to high-end translation memories and automatic-translation systems.
Two technologies are central to our discussion. Machine translation (MT) is
typically the use of computers to perform first-pass translation. It is accompanied by the building of user-specific dictionaries, and is followed by
post-editing by language professionals to verify accuracy and correct the
translation. Translation memory (TM) is a translation tool that stores previously translated material and leverages it for new translations. Depending on
the configuration and sophistication of the system, the TM solution may
have varying degrees of sophistication in its fuzzy matching capability in
relating new text to previously translated materials.
This case study reports how the integration of these two translation tools
enabled both the process and the translators at Baan Company.

204 Combining Machine Translation with Translation Memory at Baan

B A C K G R O U N D

In 1997, having evaluated various MT systems, Baan Company decided to

run a six-month MT-pilot project in Hanover, Germany, for English-intoGerman translation using the Logos MT system. Baan had already considered MT in 1994, but decided against it, based on these factors:

A substantial investment would be required in manpower and technical

equipment.

New writing guidelines would be needed for source and target texts.

Terminology would have to be used consistently to allow efficient

machine translation.

Integration of translation memory (reuse of previously translated

material) was not straightforward.

The return on investment was not clear.

While Baan postponed introduction of MT in 1994, the company began to

pave the way for introduction of the technology. They streamlined terminology, recruited more employees for the documentation and translation
department, and introduced Standard Baan English (a simplified and heavily
rule-based subset of English that minimizes ambiguitiesmaking machine
translation much more accurate and efficient).
Because English and German are often among the most developed language
modules in the MT industry and because English is Baans source language
and Germany one of Baans key markets, Baan decided that the English
into-German translation team would pilot the use of MT in the
organization.
By 1997 Baan was ready to implement MT. Management strongly supported the initiative and asked Carmen Andrs Lange (an author of this
study) to lead the pilot project and integrate MT into the English-German
translation procedure. Management established the following objective:
reduce throughput time for translations by 50 percent through the use of
automated translation.
Pi lot Pha se 1 : Tex t Analy sis an d Logos a s a S t and - Alon e Sy stem

Linguistic analysis demonstrated that Baans standard online-help texts were

the most suitable candidate for MT.
Baan obtained the Logos machine-translation package, and introduced
terms, semantic rules, and pattern-matching rules to obtain optimal output.
Interestingly, testing revealed that Baans standards for formatting text

Carmen Andrs Lange & Winfield Scott Bennett 205

proved problematic when those texts were submitted for machine translation. Written with editing software integrated into Baans development
system, the texts contained hard line breaks and numerous integrated codes.
The codes represented either hyperlink-specific software components
described in the online help, or phrases used to speed up manual translations. Text analysis showed that Logos performance would be greatly
enhanced if Baan could protect codes from being translated, and if hard line
breaks could be removed where needed. This was accomplished manually
during the pilot phase. Something as apparently trivial as removing hard
line breaks proved problematic, and close collaboration between Baan and
Logos was ultimately needed to resolve the issue.
Since Baan sought to implement a highly user-friendly solution for all
involved, it was determined at the end of Pilot Phase 1 not to use Logos as
a stand-alone solution, since too many issues remained open, not to mention challenges of internal buy-in for such a dramatic departure from
Baans traditional approach to translation.
P i lot P ha s e 2 : L ogos C o mb in ed w i t h Tra ns l at i o n Me m o r y ( TM )

The solution to Baans formatting and other problems (e.g., Baan had not
yet selected a post-editing tool) came unexpectedly. Why be concerned
with removing hard line breaks if Stars Transit Translation Memory program could handle them? Baan had been using Transit for English-toGerman translation since 1994. The idea was born to leverage both technologies in one solution.
This decision put the project back on track. An interface had recently been
developed by Star and Logos, integrating Transit Translation Memory and
Logos Machine Translation. Baan took a fresh look at the problem. During
this phase, it was necessary to define clearly which system would carry out
which tasks. It soon became obvious that Transit could support the following tasks:

protect strings so that they are not translated or destroyed by the Logos
system during translation processing;

segment titles, subtitles, sentences, listings, etc.;

automatically find and apply perfect (100-percent) matches embedded

in previously translated reference material;

leverage the companys core terminology already defined in TermStar,

the terminological database integrated in Transit (the TermStar export
profile allows Baan to export their company-specific terminology and
import it into Alex, the Logos dictionary-management package);

206 Combining Machine Translation with Translation Memory at Baan

extract completely new segments and fuzzy matches, which are

subsequently sent to Logos through the interface;

incorporate a post-editing process.

The only task left for Logos is the translation of virtually every new or modified text segment.
Pilot Phase 3: Workflow

Once Baan recognized that MT could be integrated into their translation

procedure through the TM system, they immediately began readjusting the
entire workflow. What at first seemed very complicated and novel soon
turned out to have notable similarities in structure to the prior approach of
which the steps are listed and diagrammed below:
1. Text is imported into Transit.
2. Text is checked against reference material.
3. Multiple occurrences can be written into files called extract.*
Transit Workflow Without MT
copied
document
- [Link]
- [Link]

original
document

reference
copied
material
document
- *.eng
[Link]
- *.ger
[Link]

TRANSIT

- text.*

copied
document
- [Link]
- [Link]

A dominant idea behind the new MT approach was, simply put, to create
new reference materials. The new procedure (see Figure 2) was as follows:
1. The original text (in Figure 1 called text.*) is imported into Transit.
2. Transit creates two file pairs named [Link]/[Link] and
[Link]/[Link].
3. The [Link]/[Link] file pair is compared with available reference
material. Perfect (100-percent) matches are automatically inserted in
[Link], the file for the target language German.
4. [Link] contains all untranslated segments (without redundancies) and
is sent via the interface directly from Transit to the LogosClient where

Carmen Andrs Lange & Winfield Scott Bennett 207

the user may choose the appropriate Logos translation profile. The
LogosClient sends the file to the LogosServer where the file is
translated.
5. As soon as Logos has finished the automatic translation, the translated
[Link] is, again, automatically returned to Transit where it can be
used as reference material.
6. If the translator opens the file to be translated in Transit (i.e.,
[Link]/[Link]) and activates the interactive TM function Associative
Net, he or she always obtains not only the fuzzy matches of former
translations, but also the Logos-translated segments from the [Link]
file.
7. Post-editing is done in the Associative Net window. After post-editing,
the sentence or segment is sent to the [Link] file in the target-language
window (see Figure 3).
Transit/Logos Interface
copied
document
- [Link]
- [Link]

original
document

TRANSIT

- text.*

reference
copied
document
material
- *.eng
[Link]
- [Link]
*.ger

NEW
copied
reference
document
material
--[Link]
[Link]
--[Link]
[Link]

copied
document

LOGOS

- [Link]
- [Link]

As shown in the following figure, this procedure allows human translators

to control the entire process. The TM system ensures that translators review
every segment and cross-check the terminology used by Logos with company-approved terminology in the TermStar master database.
Transit Translation Memory
source language
window
[Link]

target
language window
[Link]

terminology
database
TERMSTAR

fuzzy
matcher
reference
material
[Link]

208 Combining Machine Translation with Translation Memory at Baan

Pilot Phase 4: Enh an c ing Logos Outpu t

Baan began to enhance the procedure by adding an extra Logos translation

run, instead of immediately starting post-editing.
Translators can use the Transit Insert Extract feature to copy the content of
[Link] into [Link]. They then inspect each translated segment and forward comments to Baans Superuser (the team member authorized to enter
new terms and semantic rules into Logos). The original text is then run
through the Logos system a second time, leveraging the new information.
Baan initiated training in the new approach for each member of the
English-German team and simultaneously began to process live texts currently in the translation process. Finally, they compared the throughput
times for manual, TM-based and TM/MT-based translations.
Based on the results, shown in the figure below, at the end of the pilot
project, Baan came to several conclusions. They could reduce throughput
times by 50 to 60 percent if all ran smoothly. However, this does not mean
a procedure that runs without technical problems; it also assumes that
everyone involved fully cooperates (a significant challenge, given the traditional resistance to MT in many environments). Note that MT productivity
was markedly lower in the case of 350 text lines, actually taking longer than
manual translation.
Initial Results
hours

text lines

Th e H um an Fa c tor

Nearly every article about MT focuses either on linguistic aspects or on

how fast large amounts of texts can be handled by MT. But they rarely
explain how to motivate employees, and how to integrate MT into the
translation business procedure. Introducing MT does not imply that everybody involved is excited to be part of a highly sophisticated technical
procedure. Consider the impact on the translator: many of them are moved
from a traditional translation role into that of post-editingreviewing the

Carmen Andrs Lange & Winfield Scott Bennett 209

output produced by the machine, which can be tedious and more

mechanical.
The chart above shows one result where MT took even longer than human
translation would have taken. Analysis showed that the translator involved
was very negative towards MT, and saw the pilot project as a temporary
annoyance. Baan managers found they could generally predict the time a
translator would need for post-editing by gauging his or her attitude
towards MT.
Baan decided to move forward with MT implementation, but cautiously
asked how far to go in imposing a new workflow. Baan proceeded with
these conditions:

To keep the whole MT procedure as simple as possible, they opted to

go ahead with Transit, which each team member was comfortable with
at that point.

Everyone received training in entering terms and semantic rules in

Logos.

The role of each team member was clearly defined.

Baan noticed that motivation increased dramatically after individual training

(including training of freelancers). This human factor will, however, remain
a source of concern until professional translators are willing to use MT as a
tool to assist them in their work. Baans experience was that some finally
overcame their reservations, while others were resistant and gained nothing
from the system.
Starting in early 1998, Baan implemented an integrated TM/MT approach
into its regular workflow, using the system for current and future onlinehelp texts.
D E T A I L S

This section outlines a number of the issues and their solutions discovered
in the course of the pilot project.
Baan online-help texts mainly contain three types of sentences, presented in
the table below (strings in italicsred in the online versionare recognized dictionary entries in the Transit TM system; those in boldblue in
the online versionare protected tags in Transit TM system; question
marks in the text are inserted by the MT system to mark phrases not in its
dictionary and which it translated literally):

210 Combining Machine Translation with Translation Memory at Baan

Sentences without encoding that must be protected from being

overwritten during the MT run (1a1c).

Sentences with nonencoded software components (program names,

commands, field names, etc.) (2a).

Sentences with encoded strings for software components protected

from being translated and highlighted in bold (3a3c).
Category I: Plain Text Without Hypertext Encoding

1a

1b

1c

English

For example, if the ordered quantity is large, machines with higher

production capacities are used.

Logos
Output in
German

Zum Beispiel wenn die bestellte Menge gro ist, werden Maschinen
mit hheren Produktionskapazitten verwendet.

Desired
Output

Wenn die Bestellmenge zum Beispiel gro ist, werden Maschinen mit
hheren Produktionskapazitten verwendet.

Possible
Human
Translation

Ist die Bestellmenge beispielsweise sehr hoch, werden Maschinen mit

hheren Produktionskapazitten eingesetzt.

English

Use this session in order to copy project data to a new or existing project.

Logos
Output in
German

Verwenden Sie dieses Programm, um Projektdaten in ein neues oder

bestehendes Projekt zu kopieren.

Desired
Output

no modification necessary -

Possible
Human
Translation

Mit diesem Programm knnen Sie Projektdaten in ein neues oder

bestehendes Projekt kopieren.

English

You must specify the ledger account to which the transaction amount must
be posted.

Logos
Output in
German

Sie mssen das Sachkonto angeben, auf das der Buchungsbetrag

gebucht werden mu.

Desired
Output

no modification necessary -

Possible
Human
Translation

Geben Sie das Sachkonto an, auf das der Betrag gebucht werden soll.

Carmen Andrs Lange & Winfield Scott Bennett 211

Category II: Plain Text with Nonencoded Software Components

2a

English

You must select multiple records, then use the Assign Automatic
command.

Logos Output
in German

Sie mssen verschiedene Datenstze auswhlen, dann verwenden

Sie den Zuweisen Automatischen Befehl.

Desired
Output

Sie mssen verschiedene Datenstze auswhlen, dann verwenden

Sie den Befehl 'Automatisch zuordnen'.

Possible
Human
Translation

Whlen Sie zunchst die gewnschten Datenstze aus, und fhren

Sie anschlieend den Befehl 'Automatisch zuordnen' aus.

Category III: Plain Text with Encoded Software Components (Hyperlinks)

3a

3b

English

Select the check box %TF@[Link] in the

%SE@tiipd0501m000 session.

Logos Output
in German

Markieren Sie das Kontrollkstchen%TF@[Link]

in%SE@tiipd0501m000 Programm.

Desired
Output

Markieren Sie das Kontrollkstchen %TF@[Link] im

Programm %SE@tiipd0501m000.

Possible
Human
Translation

Markieren Sie das Kontrollkstchen %TF@[Link] im

Programm %SE@tiipd0501m000.

English

Use this session in order to enter the default project

%HC(tiGL000053,calculation office) by the project
%HC([Link],series) that is used for automatic numbering.

Logos Output
in German

Verwenden Sie dieses Programm, um das Standardprojekt

einzugeben%HC(tiGL000053,Abteilung fr Auftragsbearbeitung)
vom Projekt%HC([Link],Nummernkreis), das fr
automatische Numerierung benutzt wird.

Desired
Output

Verwenden Sie dieses Programm, um das Standardprojekt fr die

%HC(tiGL000053,Abteilung fr Auftragsbearbeitung) pro
Projekt%HC([Link],nummernkreis), der fr die
automatische Numerierung benutzt wird, einzugeben.

212 Combining Machine Translation with Translation Memory at Baan

Category III: Plain Text with Encoded Software Components (Hyperlinks)

3c

English

%FF@[Link] field is a combination of

degressive depreciation, where the depreciation base is the
%HC([Link],book value), and
%[Link], where the depreciation base is
%HC([Link],purchase value).

Logos Output
in German

%FF@[Link] das Feld ist eine Kombination

aus ?degressive Abschreibung, wo die Abschreibungsart
ist%HC([Link],Buchwert)
und%[Link] die Abschreibungsart
ist%HC([Link],Anschaffungs- und
Herstellungskosten).

Desired
Output

Das %FF@[Link] Feld ist eine Kombination

aus degressiver Abschreibung, wobei die
Abschreibungsbemessungsgrundlage der
%HC([Link],Buchwert) ist, und die
%[Link], wobei die
Abschreibungsbemessungsgrundlage die
%HC([Link],Anschaffungs- und
Herstellungskosten) sind.

The human translator can more or less translate around the codes, given that
translations into German in many cases do not require restructuring the
sentence. When the translator is finished, all codes are replaced by the target
text components.
Highlighting terms already in the terminology database (italic in the above
examples, red in the online system) greatly speeds up the process. By simply
pressing shortcut keys, translators can copy German target words into their
documents.
Another conclusion: the greater the number of nonencoded strings in a
text, the less attractive a purely human-translation approach, given that this
is where the machine is most powerful.
The previous table shows the first Logos output without using patternmatcher rules or macros. Only a few basic terms had been introduced into
Alex, the Logos dictionary-management tool, and some semantic rules had
been defined in Semantha, the Logos tool for semantic rules.
See example 3a where select the check box = Kontrollkstchen markieren is
defined by the rule:
select (v) (tr) check box(n) = markieren (v) (tr) Kontrollkstchen (n)
where v=verb, tr=transitive, n=noun
Translators confronted with the initial Logos output commonly reject it,
stating that post-editing the sentence would entail a complete rewrite.

Carmen Andrs Lange & Winfield Scott Bennett 213

The first problem was to teach Logos the meaning of the encoded strings,
which turned out to be virtually impossible, since dictionary entries in Alex
could only be defined using alpha characters. The idea of encoding all available Baan text codes as term entries into Alex had to be abandoned. Apart
from rejecting non-alphas, Logos unfortunately also destroyed the codes
during the MT run. The solution was to define all codes as protected tags
in Transit.
This produced the desired result, although translators had to live with two
constraints: the protected strings are shifted one position to the left (see
Logos Output in German for examples 3a3c) and, more importantly, sentences could not be analyzed appropriately.
Example 3b:

the project %HC([Link],series)...

is translated by Logos with

vom Projekt%HC([Link],Nummernkreis)

instead of
pro Projekt%HC([Link],nummernkreis)...

Logos overlooks the connection between project and series because of the
protected code between them.
Example 3a may be the most typical online-help sentence, as it appears regularly only with other codes for field and program names.
Select the check box %TF@[Link] in the %SE@tiipd0501m000
session.

is translated by Logos with

Markieren Sie das Kontrollkstchen%TF@[Link]
in%SE@tiipd0501m000 Programm.

In order to receive the desired output:

Markieren Sie das Kontrollkstchen %TF@[Link] im Programm
%SE@tiipd0501m000.

the English sentence has to be rewritten before sending it to Logos.

If the sentence
Select the check box %TF@[Link] in session
%SE@tiipd0501m000.

is sent to Logos, the output becomes:

214 Combining Machine Translation with Translation Memory at Baan

Markieren Sie das Kontrollkstchen%TF@[Link] im

Programm%SE@tiipd0501m000.

To brush up the output, a macro can be run shifting the protected strings
one position to the right. Now the output becomes:
Markieren Sie das Kontrollkstchen %TF@[Link] im Programm
%SE@tiipd0501m000.

No further modifications are needed if two data manipulations are performed: the noun session must be inserted before the code (a macro can be
run beforehand) and the protected codes must be shifted to the right before
post-editing (via a macro).
To this point, we have seen that text with encoded strings turns out to be
problematic for MT. Let us turn then to text that contains only natural
(human) language.
Example 2a:

The sentence
You must select multiple records, then use the Assign Automatic command.

is translated by Logos with:

Sie mssen verschiedene Datenstze auswhlen, dann verwenden Sie den
Zuweisen Automatischen Befehl.

Logos cannot recognize Assign Automatic command as a command button

with the description Assign Automatic.
Desired output ready for post-editing could be:
Sie mssen verschiedene Datenstze auswhlen, dann verwenden Sie den
Befehl 'Automatisch zuordnen'.

Assign Automatic command can be defined in Alex as a noun with the German transfer Befehl Automatisch zuordnen. The output becomes:
Sie mssen verschiedene Datenstze auswhlen, dann verwenden Sie den
Befehl Automatisch zuordnen.

What is left is to put the command description Automatisch zuordnen in quotation marks by means of a pattern-matcher rule in Logos.
REPLACE TARGET 'Automatisch' 'zuordnen' WITH '\'Automatisch'
'zuordnen\''

Not every sentence in an online-help manual includes encoded strings or

predefined names of software components. Natural-language phrases can
sometimes be time-consuming for human translations because of their

Carmen Andrs Lange & Winfield Scott Bennett 215

complex syntax. Standard English ensures simplified structure and reduces

the length of sentences, thereby heightening MT accuracy and simplifying
post-editing.
Example 1a:

The sentence
For example, if the ordered quantity is large, machines with higher production
capacities are used.

is translated by Logos with

Zum Beispiel wenn die bestellte Menge gro ist, werden Maschinen mit
hheren Produktionskapazitten verwendet.

Two constraints are immediately obvious to the translator. ordered quantity is

not translated correctly. The desired German transfer should be Bestellmenge
instead of bestellte Menge. And the word order For example, if is not changed
in the German output.
The former problem can easily be solved by redefining the Alex entry for
ordered quantity. Changing the word order for For example is trickier:
Wenn die Bestellmenge zum Beispiel gro ist, werden Maschinen mit
hheren Produktionskapazitten verwendet.

There are clearly cases like this where it may be to cumbersome to change
word order before MT processing, or where the rules of the MT system do
not provide an ideal solution.
On the other hand, some Logos translation results require little further
modification. One often hears of the howlers produced by MT, but a
sampling of sentences in sound shape after MT such as the following shows
the potential for an automated solution.
Example 1b:

Use this session in order to copy project data to a new or existing project.

is translated with:
Verwenden Sie dieses Programm, um Projektdaten in ein neues oder
bestehendes Projekt zu kopieren.

The errorless output could be achieved by defining two semantic rules telling the system that:
to use a session should always be translated with ein Programm verwenden [use
(vtr) session (n) = verwenden (vtr) Programm (n)] and

216 Combining Machine Translation with Translation Memory at Baan

to copy sth to sth should always be translated with etwas in etwas kopieren [copy
(vtr) n (acc) to (prep) n (acc) = kopieren (vtr) n (acc) in (prep) n (acc)
where v=verb, tr=transitive, n=noun, acc=grammatical case accusative,
prep=preposition

The same goes for example 1c where no further modification is required:

You must specify the ledger account to which the transaction amount must be
posted.
Sie mssen das Sachkonto angeben, auf das der Buchungsbetrag gebucht
werden mu.
E V A L U A T I N G

T H E

R E S U L T S

Since typical times for human translations were known beforehand, we

concluded that machine translation, in combination with translation memory, was able to speed up translation throughput times by up to 50 percent,
but only if everything ran smoothly:

The grammar of the input text must be correct.

Translators must run the correct macros.

The MT system must be provided with the correct terms and semantic
rules to understand the text.

Translators are properly trained and motivated to accept the style of an

output free of grammatical faults as acceptable.

Smooth runs also assume no system crashes and no problems with rules and
terminology. In the study, the time for pre-editing, MT runs, post-editing,
and dictionary work were carefully timed, as Baan was intent on quantifying the value of any savings, whether in terms of time or money.
As a result of the pilot, Baan concluded that translators willingness was
essential to success. The linguists role changed drastically from that of an
actor in the foreground who determines word-order, to one in the background who manipulates all available functions to force MT to produce an
acceptable output.
Using trial-and-error during MT-test runs, translators discovered the
following:

Carmen Andrs Lange & Winfield Scott Bennett 217

1. Phrases like Use the session to are translated without errors, if the English
is changed to Use the session in order to prior to MT (solution: macro).
2. Encoded strings that are combined with natural words like session, field,
tab are correctly translated simply by changing the code+word-order
(solution: macro).
3. Encoded strings will not be destroyed by Logos if they are protected by
translation memory (solution: project definition with regular
expressions in TM Transit).
4. Sentences containing hard returns are not correctly parsed by Logos
since Logos only analyzes the strings between two hard returns
(solution: segmentation rules by means of regular expressions in TM
Transit).
5. Nonencoded software-component descriptions can be translated
without further modifications by defining them in Alex and writing
pattern-matcher rules.
C O N C L U S I O N

A N D

F U T U R E

P R O J E C T S

During the MT pilot, Baan experienced the doubts typical of many with an
interest in MT. For instance, at the end of the first pilot phase, Baan was disappointed and could barely believe that formatting tags could really be an
insurmountable barrier to the use of MT in their environment.
What at first glance seemed to be a stopgap-solution (i.e., going through
Transit and the Transit-Logos interface) emerged as the most suitable procedure for Baan and Baans translators. Among the first to implement the
Transit/Logos interface, Baan specifically implemented these processes:

importing terminology into Logos;

protecting strings by means of regular expressions in Transit;

solving segmentation of phrases and sentences via regular expressions in

Transit (even a missing period no longer hampers analysis);

inserting previously translated material;

integrating the post-editing environment through Transits

synchronized windows for source and target language;

motivating team members.

At present writing, the LogosServer is established at Baans subsidiary in

Hanover, Germany, where the English-into-German team works.

218 Combining Machine Translation with Translation Memory at Baan

Baan anticipates that an annual average of 1.5 million words in online-help

files will be translated from English into German using the MT system. MT
work currently involves five to six translators. The online-help files are specifically for Baans Enterprise Requirements Planning (ERP) System, an
administrative system that supports corporate-wide accounting, inventory
control, production, and sales and purchasing departments. The ERP System is targeted at markets in the US, Germany, and Japan. Some 200,000
words of online help translated using MT are now part of the documentation for the ERP System in the field.
Current and future projects include further enhancements to Standard Baan
English by means of a pre-editing tool and the introduction of more language pairs to be translated with MT. Baan currently uses Trados
Translators Workbench and IBM TranslationManager in addition to Star
Transit; Logos MT is the sole MT system currently in use.
An open issue is implementing an authoring tool to enhance source text
quality not only for MT, but also for implementing Standard Baan English.
Also under consideration are:

how to handle pre-edited English texts from former software releases

that have been modified for the latest release, since TM will consider
the modified English text as new text;

knowledge transfera perennially open issue as team members leave

and new ones join the group;

recruiting appropriate translators with experience in MT and TM;

reservations and reluctance for translators to use MT.

15

L A N G U A G E A U T O M A T I O N
C A S E S T U DY

15

Language Automation
at the European Commission
Colin Brace

In the mid-1970s, at the instigation of a far-sighted European Union official, Loll Rolling (now retired), the European Commission made its first
incursion into the world of machine translation (MT). It acquired a license
to exploit Systran, a commercial machine-translation system, and began
beefing up the systems dictionaries with EU terminology and developing
additional language pairs that the Commission required. While sharing its
origins, the Commissions system, for practical purposes, has scarcely more
than name in common with the commercial PC-based software developed
and marketed by California software company Systran.
R E V E R S E - E N G I N E E R I N G

B A B E L

A T

T H E

E C

With 11 official languages to contend with, the collective institutions at the

helm of the European Union are probably the largest producer and consumer of translation in the world today. The most familiar of these is the
European Commission, the enormous administrative wing of the EU
divided between Luxembourg and Brussels. Others include the European
Parliament, the Council, the Court of Justice, the Court of Auditors, the
Economic and Social Committee, and the European Investment Bank.
Because the EU is committed to linguistic equality among its member
states, all Union legislation needs to be translated into the 11 official languages. But it doesnt stop there. Countless calls for proposals, internal
reports, and intermediary drafts need to be translated on an ad-hoc basis,
not to mention the need for live interpreting of meetings. Moreover, as
new countries join the Union and additional languages need to be reckoned with, the translation burden expands exponentially.
The major EU institutions all have their own internal translation departments, the largest of which is the Commissions Translation Service (SdT).
With an army of some 1,500 full-time professional translators divided
between Luxembourg (one third) and Brussels (two thirds), the SdT is the
largest single translation organization in the world, producing over one million pages a year.

220 Language Automation at the European Commission

M A C H I N E

TR A N S L A T I O N

J O I N S

T H E

E C

Since the early 1990s, use of MT at the Commission has soared, primarily
due to the adoption within the organization of email, which simplifies submitting and retrieving texts, but also thanks to some judicious internal
promotion. In 1996, some 220,000 pages were run through the system,
making the Commission by far the most prolific user of MT in the world in
terms of shear volume. The Commissions Translation Service (SdT)
accounted for slightly less than a third of this volume, with the remainder
being nonlinguists in the many administrative departments.
Until now, funding for development and maintenance of Systran has been
provided by DG XIII under the Multilingual Action Plans (MLAPs), but
the success of the system within the organization placed the Commission in
a quandary. DG XIIIs raison dtre is funding research in telecommunications and language engineering, and it could no longer continue to justify
subsidizing development of the system if it had truly passed from being an
advanced research topic to a fully operational concern. So in 1995, DG
XIII announced its intention to phase out its support by the end of 1997,
and over the past two years the SdT has been contemplating the way
forward.
I N - H O U S E

M T

U S E R

S U R V E Y

One option is for the SdT to allocate funds from its own operational budget
to support the use of MT within the Commission. To determine whether
this would be appropriate, an extensive feasibility study was undertaken last
year, encompassing a user survey, practical experiments with in-house translators, an examination of legal issues, a market study, and a cost-benefit
analysis.
More than 1,500 usersboth translators and nonlinguists alikeresponded
to the survey, providing a very detailed snapshot of how machine translation
is used within the Commission. With an eye on objective information, the
SdT also surveyed a number of nonusers, people who for one reason or
another do not avail themselves of the system. The results of the survey provide a unique picture of the use of MT within this vast organization.
Among its users in the administrative departments, the vast majority turn to
MT for urgent translations that they might otherwise have sent to the SdT,
for browsing, and for preparing draft versions of documents. Within the
SdT, some translators consider that MT does not help them in their work,
or remain opposed to the use of MT on principle. But many value the systems fast turnaround times, and find its vast terminological resources and its
preservation of formatting to be important benefits. While post-editing

Colin Brace 221

machine output can be tedious, some translators, as Dorothy Senez of the

MT Help Desk wryly notes, find consolation in the systems unintentional
sense of humor.
Of course, the system is not cut from the same technical cloth in every
detail, and the quality of its translations varies greatly among the language
pairs. At the moment, the French-English, French-Spanish, French-Italian,
and English-French language pairs are considered by users to be the best.
Feedback from the practical experiments carried out by the SdT shows that
on average a translation time saving of 35 percent can be achieved, provided
a number of conditions are met. Documents have to be of the appropriate
type, post-editors should be experienced, and MT dictionaries prepared in
advance. And the actual results will still depend on the quality of the language pair in question.
The study acknowledges that SdT users and administrative users have different requirements, but the general consensus is that the primary value of MT
lies in its immediacyMT is fast. They also perceive the need for improved
linguistic coverage as well as better promotion within the Commission.
So where does this leave the SdT with regard to the future of MT within
the institution? While the exact details have yet to be hammered out, it
appears that the SdT and DG XIII have reached a happy compromise. Now
that the SdTs cost-benefit analysis has demonstrated to its satisfaction that
the MT system both directly and indirectly benefits not just the Service
itself but also the Commission as a whole, the SdT will now support the
mature, operational language pairs.
DG XIII meanwhile has agreed to continue funding the development of
other language pairs, under the famous subsidiarity principledevelopment will depend on cofinancing by the relevant Member States. In other
words, if a Finnish-English language pair is deemed a priority, the government of Finland will have to be prepared to partly underwrite the effort.
C O M M E R C I A L

A R E N A

The Commission will be issuing calls for tenders for the maintenance of the
most promising language pairs as well as for systems or services for languages not covered by the Commissions system, or for language pairs
which are of lesser quality. Keeping an eye on developments in the commercial arena, the Commission could conceivably license a language pair
not covered by Systran from a third-party developer, should such a product
become available.

222 Language Automation at the European Commission

By virtue of both its substantial internal translation requirements and its

commitment to linguistic diversity, the European Commission is in many
ways an exemplary test bed for language technology such as MT. As such, it
is in the unique position of playing the roles of both user and mover. What
lessons can be drawn from the Commissions experience by other, albeit
smaller organizations?
Euramis General Architecture
Converters

Translation Memory
Client
Email

Service
dispatcher and
integrator

Text Analysis

Client

Linguistic
Resources
Database

Terminology

Machine
Translation

I N T E G R A T I N G

M T

I N T O

WO R K F L O W

For one, the MT development team and the MT user base (the SdT in particular) have enjoyed close proximity; feedback from the latter to the former
has ensured practical results. Since few organizations can justify development of their own MT system (the Pan American Health Organization
being an exception that comes to mind), this is admittedly an exceptional
albeit pertinent factor.
In addition, the Commission has striven to integrate MT within the document flow of the organization. That means a substantial investment in the
software- engineering side of thingssuch as document format filters and
integration with emailadmittedly prosaic matters which have all too often
been given short shrift by language technologists in the past. This task has
been neither easy nor trivial (see Euramis Tools Up below).
The Commission has also expended tremendous effort building up the
Systran dictionaries. The four top-rated language pairs boasted nearly
700,000 entriesand that was before the Eurodicautom data were
imported. Currently, Systran has more than four million entries distributed
across the 16 extant language pairs. This is a lesson that applies to both small
and large MT users alike.

Colin Brace 223

Given the Commissions needs and resources, any inability to leverage this
technology would have sent out a warning message to the MT community
in general. As Dorothy Senez puts it, If we cant make it work, who can?
E U R A M I S

TO O L S

U P

One way in which Systran is expected to integrate into the ECs SdT is via
the Euramis project (short for European Advanced Multilingual Information System). Launched in 1994, and due to become semioperational for a
pilot user group in the near future, Euramis aims to provide a single client
interface to a panoply of server-based translation tools and resources for the
Commissions translators, but also other end-users of translated documents.
If it reaches its long-term goals, Euramis will very likely be the most comprehensive attempt ever to integrate 25 years of natural-language processing
and tool development into a heavy-duty workflow architecture for translators and their customers.
The key components of Euramis include a translation memory (TM) facility (presumably based on the Trados WorkbenchTrados recently
confirmed that it is providing their product to the European Union institutions for use by some 2,000 in-house translators, a terminology extraction
tool, the Systran MT facility, and a Linguistic Resources Database containing the systems complete linguistic entries, with links to Eurodicautom,
among other term bases. Users with a document to translate will send an
email request to Euramis which will automatically convert the document
into SGML format with a Unicode character set, process it using selected
TM and MT resources and email back a text with automatic candidate
translations in the target language. Translators will then post-edit the result
to deliver a final version.
Euramis Workflow
CLIENT

SERVICE DISPATCHER

MACHINE TRANSLATION
Systran Document Representation

SGMLCommandstring

APPLICATION
Function Dependent
Proprietary Text
Formats
(WordPerfect,
WinWord, Excel)

Euramis-Pivot
Format Document
Representation
SGML + Unicode

LINGUISTIC
RESOURCES
DATABASE

EURODICAUTOM

Euramis
Client Application

Exchange
Format
LRD-Format
SGML + Unicode

LRD-Format
(Local LRD)

Conversion

LRD-Transfer

Internal
Format

TRANSLATOR'S
WORKBENCH
Internal
Format

224 Language Automation at the European Commission

I N S I D E R S

O N L Y

Unfortunately, for interested parties outside the Unions institutions, the

Commissions Systran is likely to remain an attractive product but out of
reach for the foreseeable future. Access to the Commissions MT system is
restricted to users within the EU institutions, the rationale being that the
Commission should avoid distorting the competitiveness of the free market.
In any event, Systran, which originated from research at Georgetown University (Washington, DC) in the 1950s, looks poised to enjoy a rosy future
well into the 21st century. If language is the soul of a culture, then the soul
of a unified but multicultural Europe lies in its multilinguality, symbolized
imperfectly yet impressively by Systran.

225

About the Contributors

Carmen Andrs Lange received a degree in technical translation at the University of Hildesheim, Germany, where she studied Spanish and English in
combination with mechanical engineering, electronics, and applied linguistics. In 1991 she began working for Space GmbH in Hanover, Germany as
a translation engineer and technical author. Space GmbH later became
Baan Germany (1995). Andrs Lange has translated software programs, user
manuals, and product information. She became team leader for the German
translation group in 1994. In March 1997, she came to the Netherlands to
investigate and lead a pilot project in machine translation for Baan Development. Since 1997 she has worked as a translation consultant, offering terminology workshops and presentations on Baans innovations in translation
memory, machine translation, and terminology. Her current project is a 16language database to be implemented on the Internet.
Winfield Scott Bennett received a Ph.D. in Germanic linguistics from the
University of Texas at Austin in 1978. After a year on the faculty at UTAustin he began research and development on the Siemens-sponsored
Metal machine-translation project at the University of Texas. During this
period he was instrumental in the research-and-development effort to create a production machine-translation system from the existing theoretical
system developed at the University of Texas from 1959 to 1979. This system, named Metal, was marketed by Siemens beginning in 1987. Siemens
stopped funding the R&D effort in 1993, at which time Bennett moved to
Logos Corporation. During his time at Logos he has served as director of
linguistics, director of institutional and government relations, director of
marketing, and now director of strategic relations. He has been active in the
international translation community for a number of years, including holding offices in the Association for Machine Translation in the Americas and
the International Machine Translation Association, as well as membership
in the ATA. Bennett has written a number of articles and contributed to
books on machine translation and related issues.
Colin Brace is a journalist based in Amsterdam who writes regularly about
language technology. From 1991 to 1996, he published the Language Industry Monitor, a newsletter dedicated to this field.
David Brooks joined Microsoft in 1990 as the manager of transfer pricing,
and since then has held a variety of jobs with an international focus. Today
he is senior director of international product strategy for Microsoft, and his
mission is to support and coordinate the creation of international versions.

226 About the Contributors

Prior to joining Microsoft, Brooks was a management consultant with

KPMG Peat Marwick. There, he was an early advocate of activity-based
costing and worked extensively with technology companies on strategic
planning and cost analysis. Brookss first real job was with Ford Motor
Company, where he was a financial analyst in International Automotive
Operations. After graduating from Cornell University in 1969 with a
degree in English literature, he spent three years in Japan teaching English.
After returning to the US, he completed his MBA degree at the University
of Washington and passed the CPA exam several years later. Brooks lives in
Seattle with his wife and two teenage children. He is an active outdoor
enthusiast who enjoys biking, hiking, and skiing, and is also an active amateur astronomer and cosmologist. His favorite saying is, Anything that
doesnt kill me makes me a better man.
Susan Cheng is the cofounder and vice president of marketing for Idiom
Technologies. She completed her A.B. in history and science at Harvard
University. While completing her undergraduate work, Cheng was program director and community-relations director for Partners for Empowering Neighborhoods, an organization that offers English as a Second
Language classes in Boston-area housing developments. Cheng also
cofounded and directed the Health Advocacy Program at Boston Medical
Center, a volunteer program that improves access to health care and welfare
for low-income children and their families. She has also held leadership
positions at numerous benevolent organizations, including the Rockefeller
Foundation and Project New Life in Boston, a literacy program for adults.
Cheng was a recipient of the 1995 Action for Boston Community Development Award. Contact her at scheng@[Link].
Karen Combe is president of Combe Consulting, which offers consulting in
internationalization and localization. Prior to establishing her practice,
Karen worked for nearly 10 years at International Language Engineering, a
localization service provider. At ILE she led the development of each of the
companys production departments during its growth from 12 people to a
staff of over 125. As senior vice president she managed both the client-services and sales and marketing departments. Karen holds a B.A. in linguistics
from the University of California at Berkeley, and a post-graduate degree in
social anthropology from Cambridge University. She spent two years in the
Peace Corps in Senegal, West Africa, and two years with International Voluntary Services in Algeria. Subsequent work experience includes two years
in the financial-services industry and eight years working as a cowboy on a
ranch in northwestern Colorado.

About the Contributors 227

Cornelia Hofmann has been working in technical communications for over

10 years. She started as technical writer in 1987 with AEG (now Schneider
Automation) in Seligenstadt, Germany, and became manager of technical
communications in 1993. Since then, her team has been working on multilingual information management aimed at improving the processing of
information in multiple languages. In 1993, a first single-source publishing
project started. Her team developed a tool based on Interleaf which allows
the creation of paper documentation and help files from a single source.
Currently, this tool is used to create more than 7,000 pages of documentation in multiple languages each year. In 1997, her team started a new singlesource publishing project based on SGML. Since mid 1999, she is partner
and managing director of Ovidius Gesellschaft fr Technische Kommunikation mbH in Hanau, Germany. Ovidius offers translation and documentation services as well as consulting for international companies, with focus
on the implementation of information management projects.
Gary Jaekel is currently managing director of TNC, the Swedish Centre for
Technical Terminology, after having served on the board of that organization for several years. From 1987 to 1997, Gary was manager, then managing director, of Ericsson Language Services, a wholly owned Ericsson
subsidiary. Language work was subsequently outsourced to Interverbum, a
language-services company, at the end of 1997, where Jaekel was responsible for vendor relations. Jaekel was also a founding member and former
president (1990-1995) of the Swedish Association of Professional Translators (SF). Gary is also a member of the Steering Committee for Language
Technology in Sweden, and has been a member of the Regional Centre
Europe steering committee under the auspices of FIT. Jaekel writes and
presents frequently on the areas of translation, localization, terminology, and
language training.
Andrew Joscelyne has been closely involved in promoting the emerging language industry for the past decade. After an academic career in linguistics in
Europe and North Africa, he began to explore the impact of the digital revolution on language and business in Europe in the mid 1980s. He helped
pioneer the reporting of language technology, taught technology awareness
to trainee language professionals, and set up a translation and communications consultancy in Paris, where he now lives. He has reported extensively
on language-industry issues for a number of publications ranging from Intermedia to Wired, including a column in Language International. As an associate
of the UK-based Equipe Consultancy, he is currently working on a number
of European Commission projects for the multilingual information society.

228 About the Contributors

J.H.G. (Harold) Kinds, Certified Public Accountant (1960) has worked in

accountancy for the last 20 years. At present he is head of the Technical
Department at Van Zwol Wijntjes, Accountants in Amersfoort, the Netherlands and the Cooperating Certified Public Accountants in Bunnik, the
Netherlands. He teaches auditing at the Catholic University in Tilburg
(Netherlands) and is a member of the Statistical Auditing Committee at the
Interuniversity Limperg Institute.
Bernhard Kohlmeier studied technology and cartography, earning a degree in
1985 from the Fachoberschule fr Technik in Munich, Germany. He
embarked on a career in electronic data-processing by starting his own
company when IBM introduced its first personal computers. In 1989 he
joined Software Products International GmbH, the German subsidiary of
SPI, the San Diego-based developer and publisher of Open Access, an early,
DOS-based application that integrated database, spreadsheet, business
graphics, word processing, and communications applications. In 1992,
Kohlmeier joined Gesellschaft fr Technische Information mbH (Gecap), a
pioneer in Germany in computer-aided word and graphics processing.
Kohlmeier eventually became director of localization at Gecap. As vice
president of strategy and planning, Kohlmeier guided localization strategies
for Gecaps subsidiaries and customers, including the first German version
of the Microsoft Encarta Encyclopedia, as well as other Microsoft products
such as Office 97. Kohlmeier now lives in Seattle where he serves as executive vice president for strategic development at Bowne Global Solutions.
S.L. Koo, an aeronautical engineer (1965), has worked in localization for
the last 15 years. At present she is director of L&L Gebruikersinformatie
B.V., a localization company based in the Netherlands. She is a member of
the Executive Committee of the Localisation Industry Standards Association (LISA).
Rose Lockwood is a founder and managing director of Equipe Consortium
Ltd., based in Cambridge, England. After over 15 years as a management
consultant (with Booz-Allen and Butler-Cox in New York) and market
analyst (with Ovum in London), she established Equipe in 1996 to offer
specialized services in the language industries. One of the original analysts
of the office-automation market in the 1980s, she has published market
studies covering a wide range of IT and software topics, including networking, groupware, messaging, and document management. Her Globalization
report, published by Ovum, was the first comprehensive study of the localization market. Since 1991 Lockwood has carried out around 30 different
studies of various aspects of the market for language technology and services, for suppliers, users, investors, and the public sector. She is currently

About the Contributors 229

engaged in assessing the impact of language in the digital-content industries, and in global e-commerce.
Thorsten Mehnert studied computer science and business administration in
Hamburg, Baton Rouge, and Karlsruhe and holds a diploma ([Link].) from the University of Karlsruhe. After graduation, he joined
Gemini Consulting where he worked as a management consultant on several reengineering projects for clients in telecommunications, trade, and
public administration. Since 1996, Mehnerts independent consulting work
has focused on information and process management in the documentation
and translation industry. As a senior information architect at Tanner Documents, he currently supports clients in optimizing the creation, management, and distribution of technical information often hidden in
publications such as pre-sales and spare part catalogs, user manuals, technical
data sheets, etc. Thorstens goal is to increase the impact of technical information on a companys competitiveness by synchronizing information engineering with the product life cycle and EDM/ERP systems, by utilizing
object-oriented and SGML/XML-based methods, and by integrating technical information into corporate information management. He can be contacted at mehnert@[Link] or [Link]@[Link].
Ricky P. Thibodeau joined MapInfo Corporation of Troy, New York in 1995
and has worked there as software localization engineer, quality assurance
supervisor, and international localization manager. He currently serves as
international product release director at MapInfo, where he oversees localization in 20 languages. Previously, he worked in the information technology department for the New York State Department of Social Services, and
network operations for the Department of Taxation and Finance. Thibodeau holds a bachelor of science degree in computer science and associate degree in data processing from the State University of New York. He
lives in Albany, New York with his wife and has his roots in Quebec. Contact him at rthibodeau@[Link].
Suzanne Topping is the owner of Localization Unlimited, which provides
localization consulting and training services. Prior to starting her own business in October 1998, Suzanne spent 10 years with Eastman Kodak, where
she founded two localization groups, which managed localization for a
range of products including thermal printers, digital printing kiosk systems,
inkjet printers, digital scanners, and digital cameras. At Kodak, Suzanne
focused on localization process improvement and on educating project
teams about localization. She began an Internationalization Special Interest
Group, authored a Localization Quick Reference Guide, developed content
for a corporate Internationalization Web site, and coordinated the creation

230 About the Contributors

of development guidelines for easily localizable software. She is currently

the localization editor for the Open Directory Project ([Link] and is founder and moderator for the Northeast Localization Special Interest Group (nelocsig@[Link]). Contact
her at stopping@[Link].
Alberto Vourvoulias-Bush is deputy editor for Time Latin America. Before
entering journalism he was a research associate at the Council on Foreign
Relations, a public policy think-tank, and a lecturer on Latin American
politics at Yale University and New York University. He has published in
both Latin America and the United States, in English and Spanish, and is
coeditor of The City and the World: New Yorks Global Future (Council on
Foreign Relations Press, 1997).
A B O U T

T H E

E D I T O R S

Robert C. Sprung is chairman and founder of Harvard Translations, Inc., a

Cambridge, Massachusetts-based foreign-language translation solutions provider and consulting firm. Sprung is editor of Language International, a trade
magazine for the language professions (translation, localization, globalization). He writes and speaks regularly on language and translation, including
engagements at the computer show Comdex, a column on language issues
in the magazine Pharmaceutical and Medical Packaging News, and a chapter in
Designing User Interfaces for International Use from Elsevier. Sprung holds an
A.B. from Harvard College, summa cum laude, in classics and German.
He received an M.A. in modern languages from the University of Cambridge (England) with first-class honors, under a George C. Marshall
scholarship. Contact him at rsprung@[Link]; [Link],
[Link].
Simone Jaroniec holds a bachelor of arts degree in French and international
relations from Kent State University. She was awarded an internship at the
Center for Human Rights of the United Nations in Geneva. She holds a
master of arts in German translation from Kent State University, where she
concentrated on terminology management. Jaroniec is currently manager of
linguistic services at Harvard Translations, Inc., where she also oversees
research and application in language tools and technologies.

231

About the Sponsors

A M E R I C A N TR A N S L A T O R S A S S O C I A T I O N
O F F I C E R S

A N D

B O A R D

Ann G. Macfarlane, President

Thomas L. West III, President-Elect
Courtney Searls-Ridge, Secretary
Eric Norman McMillan, Treasurer

R E C I P I E N T S

1964
1965
1966
1967
1968
1969
1970
1971
1972
1973
1974
1975
1977
1978
1980

O F

O F

T H E

A L E X A N D E R

Alexander Gode
Kurt Gingold
Richard and Clara Winston
The National Translations Center
(University of Texas)
Pierre-Franois Caill
Henry Fischbach
Carl V. Bertsche
Lewis Bertrand
Lewis Galantire
Jean-Paul Vinay
Eliot F. Beach
Frederick Ungar
Eugene A. Nida
Royal L. Tensley, Jr.
Gregory Rabassa

P R E S I D E N T S

1995-97 Peter W. Krawutschke

1993-95 Edith F. Losa
1991-93 Leslie Willson
1989-91 Deanna L. Hammond
1987-89 Karl Kummer
1985-87 Patricia E. Newman
1983-85 Virginia Eva Berry
1981-83 Ben Teague
1979-81 Thomas R. Bauman
1977-79 Josephine Thornton
1975-77 Royal L. Tinsley

2 0 0 0

Allan W. Adams, Kirk Anderson, Beatriz A.

Bonnet, Scott Brennan, Gertrud Graubart
Champe, Jo Anne Engelbert, Alan K.
Melby, Izumi Suzuki, Timothy Yuan

1981

1983
1984
1985
1986
1987
1988
1990
1992
1993
1996
1997

P A S T

D I R E C T O R S ,

( ATA )

O F

T H E

G O D E

M E D A L

Georgetown University, Monterey

Institute of International Studies,
State University of New York at
Binghamton
Franoise Estac
Charles M. Stern
Ludmilla Callaham
Richard Ernst
William I. Bertsche
Patricia E. Newman
Marilyn Gaddis Rose
Ben Teague
Deanna L. Hammond
Karl Kummer
Javier Collazo
William Gladstone
Danica Seleskovitch

ATA

1973-75 William I. Bertsche

1971-73 Thomas Wilds
1970-71 William I. Bertsche
(completed Moynihans term)
1969-70 Daniel Moynihan
(resigned in June 1970)
1967-69 Boris Anzlowar
1965-67 Henry Fischbach
1963-65 Kurt Gingold
1960-63 Alexander Gode

232 About the Sponsors

ATA

C O R P O R A T E

M E M B E R S

A2Z Printing Center

A L Madrid & Associates
A & M Logos International, Inc.
Able International, Inc.
Academy of Languages Translation &
Interpretation Services
Academy of Legal and Technical
Translation, Ltd.
Academy Translations
Accent Typography & Translation
Accento, The Language Company
Accents
Access Language Experts
Accu Trans, Inc.
Accura International Translations
Accurapid Translation Services, Inc.
Accurate Spanish Translations by Spanish
Business Services, Inc.
Accurate Translation, Inc.
Accuword International, Inc./dba inlingua
International
Ace Translation Center
Acentos, Marketing & Advertising &
Translations
Adams Translation Services
Adaptive Language Resources, Inc.
AD-EX Worldwide
Advance Language Studios
AE Inc. - Translations
Agnew Tech-II
AIM Translations, Inc.
Albanian Translation Services
Albors and Associates, Inc.
Alexandria Translations
Allen Translation Service
Allied Languages Cooperative
Ambassador Translating, Inc.
America Translating Services
Amway Corporation
Andalex International, Inc.
Antiquariat Literary Services, Inc.
Arabic Scientific Alliance
Argo Translation, Inc.
ASET International Services Corporation
Asian Translations, Inc.
ASIST Translation Services, Inc.

( 1

M A R C H

1 9 9 8 )

Astratec Tradues Tcnicas Ltda

ATG Language Solutions
ATL Ultrasound
AT&T Language Line Services
Auerbach International, Inc. dba
Translations Express
Avant Page
Babel, Inc.
Babel Translation Services
Baker & McKenzie
Banta Information Services Group
BCBR - Business Communications Brazil
Benemann Translation Center - BTC
Berkeley Scientific Translation Service, Inc.
Berlitz Interpretation Services
Berlitz Translation Services
Bilingual Services
Bowne Translation Services
Bradson Corporation
Bureau of Translation Services, Inc.
Burg Translation Bureau
C. P. Language Institute
CACI Language Center
Calvin International Communications, Inc.
Cambridge Translation Resources
Canadian Union of Professional &
Technical Employees
Carioni & Associates, Inc.
Carolina Polyglot, Inc.
Caterpillar, Inc.
Center for Professional Advancement/
The Language Center
Chicago Multi-Lingua Graphics, Inc.
Cial Lingua Service International
Ciba Corning Diagnostics Corporation
CinciLingua, Inc.
CITI-Translation Center, Inc.
Cogtec Corporation
Columbia Language Services
ComNet International
Computain, Inc.
Contact International
Continental Communications Agency
Copper Translation Service
CopyGroup, Inc.

About the Sponsors 233

Corporate Language Services, Inc.

Corporate Translations
Corporate Translations, Inc.
Corporate Translation Services, Inc.
Cosmopolitan Translation Bureau
Coto Interpreting, Translating & Graphics
Course Crafters, Inc.
Crestec (UK) Ltd.
Crimson Language Services
Crossword Translation Services
Cybertec, Inc.
CyraCom International, Inc.
Czech Translation Services
Delta Translation International
Die Presse Editorial, Ltd.
Digital Publishing, Inc.
Diplomatic Language Services, Inc.
Direct Language Communications
Diversified Language Institute
Documents International, Inc.
DocuTrans
Dynamic Language Center, Ltd.
East-West Concepts, Inc.
Echo International
Edimax, S.A. de C.V.
Elite Language Productions
Elucidex
Eriksen Translations Inc.
Escalante Translations
Euro - Translation
Excel Translations, Inc.
Executive Linguist Agency, Inc.
Expert Language Services
First Translation Services
FLS, Inc.
Foreign Ink Ltd.
Foreign Language Center
Galaxy Systems, Inc.
Garjak International, Inc.
Geonexus Communications
Geotext Translations
GeoText Translations & Typesetting
Services
Global Advanced Translation Services Inc.
The Global Institute of Languages and
Culture, Inc.

Global Language Services, Inc.

Global Language Solutions/
The Russian Word, Inc.
Global Languages & Cultures, Inc.
Global Translation Services, Inc.
Global Translations & Interpreters
Services, Inc.
The Global Word, Inc.
GlobalDoc, Inc.
Globalink, Inc.
Glorbet Consultants, Inc.
Harvard Translations, Inc.
Health Outcomes Group
Heitmann of America, Inc.
HG Translations
Hightech Passport Limited
Honda R&D North America, Inc.
HSN Linguistic Services Ltd.
i. b. d., Ltd.
IBS-International Business Services
ICN Language Services
Idem Translation
in French only inc./in Spanish too!
Information Builders, Inc.
Inlingua International Services
Inlingua Language & Intercultural Services
Inlingua Language Services Center
Institut fr Fremdsprachen und
Auslandskunde
Intel Corporation
Interclub, Inc.
Intercontact - Peru
Interlanguage SNC Di Abbati A. E C.
InterNation Inc.
International Communication by
Design, Inc.
International Communications, Inc.
International Contact, Inc.
International Effectiveness Centers
International Language Engineering Corp.
International Language Services, Inc.
International Translation and
Publishing, Ltd.
International Translators International &
Typesetters, Inc.
International Access/Ability Corp.
Interpretations

234 About the Sponsors

Interpreters International & Translations
Interpreters Unlimited
Interpreting Services International Inc.
InterSol, Inc.
Interspeak Translations, Inc.
Intertech Translations, Ltd.
InterTrans, Inc.
Inwords, Inc.
IRU International Resources Unlimited
Iverson Language Associates, Inc.
J.D. Edwards & Company, Inc.
Jackson Graphics, Inc.
Japan - America Management, Ltd.
Japanese Language Services, Inc.
JKW International, Inc.
JLS Language Corporation
John Benjamins Publishing Company
Josef Silny & Associates, Inc.
JTG, Inc.
K & L Language Services
Korean Technical Communications
Langua Translations, Inc.
Language Company Translations, L.C.
The Language Connection
The Language Exchange, Inc.
Language Innovations, LLC
Language Intelligence, Ltd.
Language Interface, Ltd.
The Language Lab
Language Link Corporation
Language Management International
(The Corporate World, Inc.)
Language Matters
The Language Network, Inc.
Language Plus
The Language Service, Inc.
Language Services Associates
The Language Solution, Inc.
Languages International
The Languageworks, Inc.
Latin American Translators Network, Inc.
Legal Interpreting Services, Inc.
Liaison Language Center
Liaison Multilingual Services
Lingo Systems
Lingua Communications Translation
Services
Linguae Translation & Interpretation
Bureau

Lingualink Incorporated
LinguaNet, Inc.
Linguistic Consulting Enterprises, Inc
Linguistic Systems, Inc.
Localization Associates of Utah
Logos Corporation
LRA Interpreters, Inc.
Lucent Technologies-ILT Solutions
Luz
M2 Limited
Magnus International Trade Services Corp.
Master Translating Services, Inc.
MasterWord Services, Inc.
Gene Mayer Associates
McDonalds Corporation
Ralph McElroy Translation Company
McNeil Technologies, Inc.
ME Sharpe, Inc., Publisher
Mercury Marine
Metropolitan Interpreters & Translators
Worldwide, Inc.
Mitaka Limited
Morales Dimmick Translation Service Inc.
Morgan Guaranty Trust Company
Multilingual Translations, Inc.
N.O.W. Translations
NCS Enterprises, Inc.
New England Translations
Newtype, Inc.
NIS International Services
Ntext Translations
Occidental Oil & Gas Corporation
Okada & Sellin Translations, LLC
Omega International
OmniLingua, Inc.
Oriental Communication Services, Inc.
OSullivan Menu Corporation
Pacific Interpreters, Inc.
Pacific Ring Services, Inc.
Paragon Language Services
Peritus Precision Translations, Inc.
Peters Translation, Inc.
Planning S.N.C.
Polyglot International
Precision Translating Services, Inc.
Premier Translation Services, Ltd.
Prisma International
Professional Translating Services
ProTrans, Inc.

About the Sponsors 235

PSC, Inc.
Quantum, Inc.
Quark, Inc.
Quintana Multi-Lingual Services, Inc.
R.R. Donnelley Financial Translation
Services
Rapport International
Rennert Bilingual Translations
Resource Network International, Inc.
Richard Schneider Enterprises, Inc.
Rosetta
Routledge, Inc.
RussTech
Sally Low & Associates
Schreiber Translations, Inc.
SH3, Inc.
Shoreline Translations, Inc.
Showorks, Inc.
Simulacrum LLC/Context
SinoMetrics International, Inc.
Slovak Translation Services
Sohsei, Inc.
Spectrum Multilanguage Communications
Sputnik Translation Services
Square D Company
Suzuki, Myers & Associates, Ltd.
Sykes Enterprises, Incorporated
Tech Link, Inc.
Techlingua, Inc.
Technik-Sprachendienst Gmbh
Techno-Graphics & Translations, Inc.
TechTrans International, Inc.
Techworld Language Services, Inc.
Terra Pacific Writing Corporation
TEXTnology CORPoration
TIICO-Translating Interpreting
International Company
Total Benefit Communications, Inc.
Trade Typographers, Inc.
Trados Corporation
Traducciones LinguaCorp
TransACT
Trans-Caribe Communications
Transcript Communications, Inc.
Transemantics, Inc.
Transglobal Translations & Immigration
Services, Inc.
TransImage

Translation Company of America, Inc.

Translation Services International Inc.
TranslationPlus
Translingua, Inc.
TransLingual, Ltd.
Transperfect
Transperfect Translations International, Inc.
Universal Translations, Inc.
Universe Interpreters and Translators
Corporation
University Language Center, Inc.
U.S. Technical Translations, Inc.
U.S. Translation Company
Vanguard Academy
Victory Productions, Inc.
Vormbrock Translating, Inc.
West-Star Consultants
Whitman Language Services
Winter Wyman Contract Services, Inc.
WKI International Communications
Wordnet, Inc.
World Trade Center Portland
Worldwide Translations
YAR Communications, Inc.

236 About the Sponsors

ATA I N S T I T U T I O N A L M E M B E R S ( 1 M A R C H 1 9 9 8 )

Academy Interpreting and Translations

International
American Education Research Corp.
American Institute of Physics
An-Nahda Educational Office
Binghamton University (SUNY)
Boston School of Modern Languages, Inc.
California State University
Center for Applied Linguistics
Chitco
Christian Science Publishing Society
The Church of Jesus Christ of the Latterday Saints
Community Interpreter Services Catholic
Charities/Greater Boston
Community Management Staff
Eureka - Foreign College Evaluators &
Translators
Executive Office for Immigration Review
FBIS
Florida A&M University
The French Library & Cultural Center
Gallaudet University
Georgetown University
Georgia State University
Instituto Superior de Interpretes y
Traductores, S. C.
Inter-American Development Bank (IDB)
Inter-American Air Forces Academy
International Refugee Center of OregonInternational Language Bank
International Institute of Connecticut, Inc.
International Institute, Inc.
Kent State University
Language Interpreter Services and
Translations/WA State Department of
Social and Health Services

The Language School

M. D. Anderson Cancer Center
Marygrove College
Mayo Medical Center
Monterey Institute of International Studies
New York University
Northern Illinois University
Ordre des Traducteurs et Interprtes Agrs
du Qubec
Purdue University
Quba Institute of Arabic and Islamic Studies
Queen of the Valley Hospital
Rose-Hulman Institute of Technology
San Diego City Schools
School District of Palm Beach County
International Student
Support/Multicultural Awareness
Department
Southwest Washington Medical Center
Summer Institute of Linguistics
Thammasat University
Translation & Critical Languages Institute
of Florida A & M Univ.
Tucson Unified School District Title I
Program/School-Community
Relations Dept.
University of Hawaii
University of Idaho
University of Miami
University of Nebraska at Kearney
University of La Verne
The University of Texas Medical Branch at
Galveston
University of Washington
Western Michigan University
Western Wisconsin Technical College
World Bank

237

For Further Reference

O R G A N I Z A T I O N S

American Translators Association (ATA). [Link]

Localisation Industry Standards Association (LISA). [Link]
Society for Technical Communication (STC). [Link]
L A N G U A G E - I N D U S T R Y

P U B L I C A T I O N S

Language International. Amsterdam: John Benjamins.

[Link]
Multilingual Computing & Technology. Sandpoint, Idaho:
Multilingual Computing. [Link]
Technical Communication, Journal of the Society for Technical Communication.
[Link]
C R O S S - C U L T U R A L

A D A P T A T I O N

Axtell, R. Dos and Taboos of Using English Around the World. New York:
John Wiley, 1995.
. Gestures: The Dos and Taboos of Body Language Around the
World. New York: John Wiley, 1991.
The Parker Pen Company. Dos and Taboos Around the World.
Elmsford, New York: The Benjamin Company, 1985.
Usunier, J. Marketing Across Cultures. New York: Prentice Hall PTR, 1996.
G L O B A L

C O M M U N I C A T I O N S

Barrett, E., ed. Text, ConText and HyperText: Writing with and for the
Computer. Cambridge, Massachusetts: MIT Press, 1988.
Brusaw, C., et al. Handbook of Technical Writing, 2nd ed. New York:
St. Martins Press, 1982.
Burnett, R. Technical Communication, 3rd ed. Belmont, California:
Wadsworth Publishing Company, 1994.
Hoft, N. International Technical Communication: How to Export Information
about High Technology. New York: John Wiley & Sons, 1995.
Leki, I. The Technical Editor and the Non-Native Speaker of English.
Technical Communication Vol. 37 No. 2 (1990): pp. 148-152.

238 For Further Reference

Sanderlin, S. Preparing Instruction Manuals for Non-English Readers.

Technical Communication Vol. 35 No. 2 (1988): pp. 96-100.
Schur, N. British English, A to Zed. New York: Harper, 1991.
S O F T W A R E
A N D

L O C A L I Z A T I O N

I N T E R N A T I O N A L I Z A T I O N

Carter, D. Writing Localizable Software for the Macintosh. Reading,

Massachusetts: Addison-Wesley, 1992.
Esselink, B. A Practical Guide to Software Localization. Amsterdam: John
Benjamins, 1998.
Hall, P. and R. Hudson. Software Without Frontiers: A Multiplatform,
Multicultural, Multinational Approach. New York: Wiley, 1997.
Jones, S., et al. Developing International User Information. Bedford,
Massachusetts: Digital Press, 1992.
Kano, N. Developing International Software for Windows 95 and Windows NT.
Redmond, Washington: Microsoft Press, 1995.
Luong, T., et al. Internationalization: Developing Software for Global Markets.
New York: John Wiley & Sons, 1995.
Nielsen, J., ed. Advances in Human Factors/Ergonomics, vol. 13: Designing User
Interfaces for International Use. Amsterdam: Elsevier, 1990.
ODonnell, S.M. Programming for the World: A Guide to Internationalization.
Englewood Cliffs, New Jersey: Prentice Hall PTR, 1994.
Taylor, D. Global Software: Developing Applications for the International Market.
Springer-Verlag, 1992.
Uren, E., et al. Software Internationalization and Localization: An Introduction.
New York: Van Nostrand Reinhold, 1993.
L A N G U A G E
A N D

TO O L S

M A C H I N E

TR A N S L A T I O N

Asian Association for Machine Translation (AAMT).

[Link]/aamt/[Link]
Association for Machine Translation in the Americas (AMTA).
[Link]/natural-language/organizations/[Link]
Center for Machine Translation and Language Technologies Institute.
[Link]

For Further Reference 239

European Association for Machine Translation (EAMT). [Link]

European Language Resources Association (ELRA).
[Link]/elra/[Link]
Gingras, B. Simplified English in Maintenance Manuals.
Technical Communication Vol. 34 No. 1 (1987): pp. 24-28.
Hearne, V. Controlled English and Poetic Common Sense.
Language Technology 9 (1989): pp. 17-20.
Machine Translation Journal. [Link]
MT News International. [Link]/eamt/mtni
Strong, K.I. Kodak International Service Language.
Technical Communication Vol. 30 No. 2 (1983): pp. 20-22.
The Unicode Consortium. [Link]
L A N G U A G E - I N D U S T R Y
A N D

M A R K E T

A N A L Y S I S

S T U D I E S

Language Translation: World Market Overview, Current Developments and

Competitive Assessment. Oyster Bay, New York: Allied Business Intelligence,
1998.
Lockwood, Rose. Globalisation 1998. Cambridge, UK: Equipe
Consortium, 1998.
. Globalisation: Creating New Markets with Translation Technology.
London: Ovum, 1995.
Mason, J. and A. Rinsche. Translation Technology Products. London:
Ovum, 1995.

American Translators Association Scholarly Monograph Series ISSN 0890 4111

I.
GADDIS ROSE, Marilyn (ed): Translation Excellence. 1987. Out of print.
II. VASCONCELLOS, Muriel (ed.): Technology as Translation Strategy. 1988.
III. KRAWUTSCHKE, Peter W. (ed.): Translator and Interpreter Training and Foreign
Language Pedagogy. 1989. Out of print.
IV. BOWEN, David and Margareta BOWEN (eds): Interpreting - Yesterday, today and
tomorrow. 1990. Out of print.
V. LARSON, Mildred L. (ed.): Translation: Theory and practice, tension and interdependence. 1991.
VI. WRIGHT, Sue Ellen and Leland D. WRIGHT Jr.: Scientific and Technical Translation. 1993.
VII. HAMMOND, Deanna L. (ed.): Professional Issues for Translators and Interpreters.
1994.
VIII. MORRIS, Marshall (ed.): Translation and the Law. 1995.
IX. LABRUM, Marian B. (ed.): The Changing Scene in World Languages. Issues and
challenges. 1997.
X. FISCHBACH, Henry (ed.): Translation and Medicine. 1998.
XI. SPRUNG, Robert (ed.): Translating Into Success: Cutting-edge strategies for going
multilingual in a global age. 2000.
All titles available from John Benjamins Publishing Company, Amsterdam/Philadelphia.
Out of print titles can be supplied as Publishing on Demand documents.