Introduction to Information Retrieval

Introduction to
Information Retrieval
CS276: Information Retrieval and Web Search
Christopher Manning and Prabhakar Raghavan

Lecture 2: The term vocabulary and postings

lists
Introduction to Information Retrieval Ch. 1

Recap of the previous lecture

 Basic inverted indexes:
 Structure: Dictionary and Postings

 Key step in construction: Sorting

 Boolean query processing
 Intersection by linear time “merging”
 Simple optimizations
 Overview of course topics
Introduction to Information Retrieval

Plan for this lecture

Elaborate basic indexing
 Preprocessing to form the term vocabulary
 Documents
 Tokenization
 What terms do we put in the index?
 Postings
 Faster merges: skip lists
 Positional postings and phrase queries
Introduction to Information Retrieval

Recall the basic indexing pipeline

Documents to Friends, Romans, countrymen.
be indexed.

Tokenizer
Token stream. Friends Romans Countrymen

Linguistic modules
Modified tokens. friend roman countryman

Indexer friend 2 4
roman 1 2
Inverted index.
countryman 13 16
Introduction to Information Retrieval Sec. 2.1

Parsing a document
 What format is it in?
 pdf/word/excel/html?
 What language is it in?
 What character set is in use?

Each of these is a classification problem, which we

will study later in the course.

But these tasks are often done heuristically …

Introduction to Information Retrieval Sec. 2.1

Complications: Format/language
 Documents being indexed can include docs from
many different languages
 A single index may have to contain terms of several
languages.
 Sometimes a document or its components can
contain multiple languages/formats
 French email with a German pdf attachment.
 What is a unit document?
 A file?
 An email? (Perhaps one of many in an mbox.)
 An email with 5 attachments?
 A group of files (PPT or LaTeX as HTML pages)
Introduction to Information Retrieval

TOKENS AND TERMS

Introduction to Information Retrieval Sec. 2.2.1

Tokenization
 Input: “Friends, Romans and Countrymen”
 Output: Tokens
 Friends
 Romans
 Countrymen
 A token is an instance of a sequence of characters
 Each such token is now a candidate for an index
entry, after further processing
 Described below
 But what are valid tokens to emit?
Introduction to Information Retrieval Sec. 2.2.1

Tokenization
 Issues in tokenization:
 Finland’s capital 
Finland? Finlands? Finland’s?
 Hewlett-Packard  Hewlett and Packard as two
tokens?
 state-of-the-art: break up hyphenated sequence.
 co-education
 lowercase, lower-case, lower case ?
 It can be effective to get the user to put in possible hyphens

 San Francisco: one token or two?

 How do you decide it is one token?
Introduction to Information Retrieval Sec. 2.2.1

Numbers
 3/20/91 Mar. 12, 199120/3/91
 55 B.C.
 B-52
 My PGP key is 324a3df234cb23e
 (800) 234-2333
 Often have embedded spaces
 Older IR systems may not index numbers
 But often very useful: think about things like looking up error
codes/stacktraces on the web
 (One answer is using n-grams: Lecture 3)
 Will often index “meta-data” separately
 Creation date, format, etc.
Introduction to Information Retrieval Sec. 2.2.1

Tokenization: language issues

 French
 L'ensemble  one token or two?
 L ? L’ ? Le ?
 Want l’ensemble to match with un ensemble
 Until at least 2003, it didn’t on Google
 Internationalization!

 German noun compounds are not segmented

 Lebensversicherungsgesellschaftsangestellter
 ‘life insurance company employee’
 German retrieval systems benefit greatly from a compound splitter
module
 Can give a 15% performance boost for German
Introduction to Information Retrieval Sec. 2.2.1

Tokenization: language issues

 Chinese and Japanese have no spaces between
words:
 莎拉波娃现在居住在美国东南部的佛罗里达。
 Not always guaranteed a unique tokenization
 Further complicated in Japanese, with multiple
alphabets intermingled
 Dates/amounts in multiple formats
フォーチュン 500 社は情報不足のため時間あた $500K( 約 6,000 万円 )

Katakana Hiragana Kanji Romaji

End-user can express query entirely in hiragana!

Introduction to Information Retrieval Sec. 2.2.1

Tokenization: language issues

 Arabic (or Hebrew) is basically written right to left, but
with certain items like numbers written left to right
 Words are separated, but letter forms within a word
form complex ligatures

 ← → ←→ ← start
 ‘Algeria achieved its independence in 1962 after 132
years of French occupation.’
 With Unicode, the surface presentation is complex, but the
stored form is straightforward
Introduction to Information Retrieval Sec. 2.2.2

Stop words
 With a stop list, you exclude from the dictionary
entirely the commonest words. Intuition:
 They have little semantic content: the, a, and, to, be
 There are a lot of them: ~30% of postings for top 30 words
 But the trend is away from doing this:
 Good compression techniques (lecture 5) means the space for
including stopwords in a system is very small
 Good query optimization techniques (lecture 7) mean you pay little
at query time for including stop words.
 You need them for:
 Phrase queries: “King of Denmark”
 Various song titles, etc.: “Let it be”, “To be or not to be”
 “Relational” queries: “flights to London”
Introduction to Information Retrieval Sec. 2.2.3

Normalization to terms
 We need to “normalize” words in indexed text as well
as query words into the same form
 We want to match U.S.A. and USA
 Result is terms: a term is a (normalized) word type,
which is an entry in our IR system dictionary
 We most commonly implicitly define equivalence
classes of terms by, e.g.,
 deleting periods to form a term
 U.S.A., USA  USA
 deleting hyphens to form a term
 anti-discriminatory, antidiscriminatory  antidiscriminatory
Introduction to Information Retrieval Sec. 2.2.3

Normalization: other languages

 Accents: e.g., French résumé vs. resume.
 Umlauts: e.g., German: Tuebingen vs. Tübingen
 Should be equivalent
 Most important criterion:
 How are your users like to write their queries for these
words?

 Even in languages that standardly have accents, users

often may not type them
 Often best to normalize to a de-accented term
 Tuebingen, Tübingen, Tubingen  Tubingen
Introduction to Information Retrieval Sec. 2.2.3

Normalization: other languages

 Normalization of things like date forms
 7 月 30 日 vs. 7/30
 Japanese use of kana vs. Chinese characters

 Tokenization and normalization may depend on the

language and so is intertwined with language
detection Is this
Morgen will ich in MIT … German “mit”?

 Crucial: Need to “normalize” indexed text as well as

query terms into the same form
Introduction to Information Retrieval Sec. 2.2.3

Case folding
 Reduce all letters to lower case
 exception: upper case in mid-sentence?
 e.g., General Motors
 Fed vs. fed
 SAIL vs. sail
 Often best to lower case everything, since
users will use lowercase regardless of
‘correct’ capitalization…

 Google example:
 Query C.A.T.
 #1 result is for “cat” (well, Lolcats) not
Caterpillar Inc.
Introduction to Information Retrieval Sec. 2.2.3

Normalization to terms

 An alternative to equivalence classing is to do

asymmetric expansion
 An example of where this may be useful
 Enter: window Search: window, windows
 Enter: windows Search: Windows, windows, window
 Enter: Windows Search: Windows
 Potentially more powerful, but less efficient
Introduction to Information Retrieval

Thesauri and soundex

 Do we handle synonyms and homonyms?
 E.g., by hand-constructed equivalence classes
 car = automobile color = colour
 We can rewrite to form equivalence-class terms
 When the document contains automobile, index it under car-
automobile (and vice-versa)
 Or we can expand a query
 When the query contains automobile, look under car as well
 What about spelling mistakes?
 One approach is soundex, which forms equivalence classes
of words based on phonetic heuristics
 More in lectures 3 and 9
Introduction to Information Retrieval Sec. 2.2.4

Lemmatization
 Reduce inflectional/variant forms to base form
 E.g.,
 am, are, is  be
 car, cars, car's, cars'  car
 the boy's cars are different colors  the boy car be
different color
 Lemmatization implies doing “proper” reduction to
dictionary headword form
Introduction to Information Retrieval Sec. 2.2.4

Stemming
 Reduce terms to their “roots” before indexing
 “Stemming” suggest crude affix chopping
 language dependent
 e.g., automate(s), automatic, automation all reduced to
automat.

for example compressed for exampl compress and

and compression are both compress ar both accept
accepted as equivalent to as equival to compress
compress.
Introduction to Information Retrieval Sec. 2.2.4

Porter’s algorithm
 Commonest algorithm for stemming English
 Results suggest it’s at least as good as other stemming
options
 Conventions + 5 phases of reductions
 phases applied sequentially
 each phase consists of a set of commands
 sample convention: Of the rules in a compound command,
select the one that applies to the longest suffix.
Introduction to Information Retrieval Sec. 2.2.4

Typical rules in Porter

 sses  ss
 ies  i
 ational  ate
 tional  tion

 Weight of word sensitive rules

 (m>1) EMENT →
 replacement → replac
 cement → cement
Introduction to Information Retrieval Sec. 2.2.4

Other stemmers
 Other stemmers exist, e.g., Lovins stemmer
 [Link]

 Single-pass, longest suffix removal (about 250 rules)

 Full morphological analysis – at most modest

benefits for retrieval
 Do stemming and other normalizations help?
 English: very mixed results. Helps recall for some queries but
harms precision on others
 E.g., operative (dentistry) ⇒ oper
 Definitely useful for Spanish, German, Finnish, …
 30% performance gains for Finnish!
Introduction to Information Retrieval Sec. 2.2.4

Language-specificity
 Many of the above features embody transformations
that are
 Language-specific and
 Often, application-specific
 These are “plug-in” addenda to the indexing process
 Both open source and commercial plug-ins are
available for handling these
Introduction to Information Retrieval Sec. 2.2

Dictionary entries – first cut

[Link]

時間 .japanese

[Link]
These may be
[Link] grouped by language
(or not…).
[Link] More on this in
[Link] ranking/query
processing.
[Link]

[Link]
Introduction to Information Retrieval

FASTER POSTINGS MERGES:

SKIP POINTERS/SKIP LISTS
Introduction to Information Retrieval Sec. 2.3

Recall basic merge

 Walk through the two postings simultaneously, in
time linear in the total number of postings entries

2 4 8 41 48 64 128 Brutus
2 8
1 2 3 8 11 17 21 31 Caesar

If the list lengths are m and n, the merge takes O(m+n)

operations.

Can we do better?
Yes (if index isn’t changing too fast).
Introduction to Information Retrieval Sec. 2.3

Augment postings with skip pointers

(at indexing time)
41 128
2 4 8 41 48 64 128

11 31
1 2 3 8 11 17 21 31

 Why?
 To skip postings that will not figure in the search
results.
 How?
 Where do we place skip pointers?
Introduction to Information Retrieval Sec. 2.3

Query processing with skip pointers

41 128
2 4 8 41 48 64 128

11 31
1 2 3 8 11 17 21 31

Suppose we’ve stepped through the lists until we process 8 on

each list. We match it and advance.

We then have 41 and 11 on the lower. 11 is smaller.

But the skip successor of 11 on the lower list is 31, so

we can skip ahead past the intervening postings.
Introduction to Information Retrieval Sec. 2.3

Where do we place skips?

 Tradeoff:
 More skips  shorter skip spans  more likely to skip.
But lots of comparisons to skip pointers.
 Fewer skips  few pointer comparison, but then long skip
spans  few successful skips.
Introduction to Information Retrieval Sec. 2.3

Placing skips
 Simple heuristic: for postings of length L, use L
evenly-spaced skip pointers.
 This ignores the distribution of query terms.
 Easy if the index is relatively static; harder if L keeps
changing because of updates.

 This definitely used to help; with modern hardware it

may not (Bahle et al. 2002) unless you’re memory-
based
 The I/O cost of loading a bigger postings list can outweigh
the gains from quicker in memory merging!
Introduction to Information Retrieval

PHRASE QUERIES AND POSITIONAL

INDEXES
Introduction to Information Retrieval Sec. 2.4

Phrase queries
 Want to be able to answer queries such as “stanford
university” – as a phrase
 Thus the sentence “I went to university at Stanford”
is not a match.
 The concept of phrase queries has proven easily
understood by users; one of the few “advanced search”
ideas that works
 Many more queries are implicit phrase queries
 For this, it no longer suffices to store only
<term : docs> entries
Introduction to Information Retrieval Sec. 2.4.1

A first attempt: Biword indexes

 Index every consecutive pair of terms in the text as a
phrase
 For example the text “Friends, Romans, Countrymen”
would generate the biwords
 friends romans
 romans countrymen
 Each of these biwords is now a dictionary term
 Two-word phrase query-processing is now
immediate.
Introduction to Information Retrieval Sec. 2.4.1

Longer phrase queries

 Longer phrases are processed as we did with wild-
cards:
 stanford university palo alto can be broken into the
Boolean query on biwords:
stanford university AND university palo AND palo alto

Without the docs, we cannot verify that the docs

matching the above Boolean query do contain the
phrase.
Can have false positives!
Introduction to Information Retrieval Sec. 2.4.1

Extended biwords
 Parse the indexed text and perform part-of-speech-tagging
(POST).
 Bucket the terms into (say) Nouns (N) and
articles/prepositions (X).
 Call any string of terms of the form NX*N an extended biword.
 Each such extended biword is now made a term in the
dictionary.
 Example: catcher in the rye
N X X N
 Query processing: parse it into N’s and X’s
 Segment query into enhanced biwords
 Look up in index: catcher rye
Introduction to Information Retrieval Sec. 2.4.1

Issues for biword indexes

 False positives, as noted before
 Index blowup due to bigger dictionary
 Infeasible for more than biwords, big even for them

 Biword indexes are not the standard solution (for all

biwords) but can be part of a compound strategy
Introduction to Information Retrieval Sec. 2.4.2

Solution 2: Positional indexes

 In the postings, store, for each term the position(s) in
which tokens of it appear:

<term, number of docs containing term;

doc1: position1, position2 … ;
doc2: position1, position2 … ;
etc.>
Introduction to Information Retrieval Sec. 2.4.2

Positional index example

<be: 993427;
1: 7, 18, 33, 72, 86, 231; Which of docs 1,2,4,5
2: 3, 149; could contain “to be
4: 17, 191, 291, 430, 434; or not to be”?
5: 363, 367, …>

 For phrase queries, we use a merge algorithm

recursively at the document level
 But we now need to deal with more than just
equality
Introduction to Information Retrieval Sec. 2.4.2

Processing a phrase query

 Extract inverted index entries for each distinct term:
to, be, or, not.
 Merge their doc:position lists to enumerate all
positions with “to be or not to be”.
 to:
 2:1,17,74,222,551; 4:8,16,190,429,433; 7:13,23,191; ...
 be:
 1:17,19; 4:17,191,291,430,434; 5:14,19,101; ...
 Same general method for proximity searches
Introduction to Information Retrieval Sec. 2.4.2

Proximity queries
 LIMIT! /3 STATUTE /3 FEDERAL /2 TORT
 Again, here, /k means “within k words of”.
 Clearly, positional indexes can be used for such
queries; biword indexes cannot.
 Exercise: Adapt the linear merge of postings to
handle proximity queries. Can you make it work for
any value of k?
 This is a little tricky to do correctly and efficiently
 See Figure 2.12 of IIR
 There’s likely to be a problem on it!
Introduction to Information Retrieval Sec. 2.4.2

Positional index size

 You can compress position values/offsets: we’ll talk
about that in lecture 5
 Nevertheless, a positional index expands postings
storage substantially
 Nevertheless, a positional index is now standardly
used because of the power and usefulness of phrase
and proximity queries … whether used explicitly or
implicitly in a ranking retrieval system.
Introduction to Information Retrieval Sec. 2.4.2

Positional index size

 Need an entry for each occurrence, not just once per
document
 Index size depends on average document size Why?
 Average web page has <1000 terms
 SEC filings, books, even some epic poems … easily 100,000
terms
 Consider a term with frequency 0.1%
Document size Postings Positional postings
1000 1 1
100,000 1 100
Introduction to Information Retrieval Sec. 2.4.2

Rules of thumb
 A positional index is 2–4 as large as a non-positional
index
 Positional index size 35–50% of volume of original
text
 Caveat: all of this holds for “English-like” languages
Introduction to Information Retrieval Sec. 2.4.3

Combination schemes
 These two approaches can be profitably
combined
 For particular phrases (“Michael Jackson”, “Britney
Spears”) it is inefficient to keep on merging positional
postings lists
 Even more so for phrases like “The Who”
 Williams et al. (2004) evaluate a more
sophisticated mixed indexing scheme
 A typical web query mixture was executed in ¼ of the
time of using just a positional index
 It required 26% more space than having a positional
index alone
Introduction to Information Retrieval

Resources for today’s lecture

 IIR 2
 MG 3.6, 4.3; MIR 7.2
 Porter’s stemmer:
[Link]
 Skip Lists theory: Pugh (1990)
 Multilevel skip lists give same O(log n) efficiency as trees
 H.E. Williams, J. Zobel, and D. Bahle. 2004. “Fast Phrase
Querying with Combined Indexes”, ACM Transactions on
Information Systems.
[Link]
 D. Bahle, H. Williams, and J. Zobel. Efficient phrase querying with an
auxiliary index. SIGIR 2002, pp. 215-221.