CSA4006: Natural

Language Processing
Winter Semester 2024-25
SCOPE
VIT Bhopal University, Sehore
Slot:
Class No.:
1
Instructor Contacts
● Instructor:
Dr. Amit Kumar
AI Researcher- Generative AI and NLP
Assistant Professor

Office: C-526
amitkumar@[Link]

2
Module-5

3
Machine Translation
❏ Machine Translation (MT) is the process of using software to
translate text or speech from one language to another
automatically.
❏ It's a key subfield of computational linguistics and artificial
intelligence.
❏ The definition may look simple, but the process is extremely
difficult. Languages differ in so many ways, grammatically,
syntactically (sentence structure), semantically (meanings), etc.
Such differences are called translation divergences.
Machine Translation
For instance, consider a quote from a Korean drama Oh my Venus,

Korean: mid-eumyeon hal su isso, hal su issdago mid nikka

English: if [you] believe [it] [you] can do [it], because [you] believe that
[you] can
Machine Translation
❏ As we can see, in terms of structure, the Korean language follows an
SOV(Subject-Object-Verb) agreement, while English follows
SVO(Subject-Verb-Object).
❏ Languages like Korean and Japanese don’t explicitly use pronouns.
❏ Such languages that omit pronouns are called pro-drop languages.
❏ These kinds of differences make machine translation difficult because
such omissions have to be identified and interpreted correctly.
❏ There are several other differences as well in these two languages.
❏ A translator must analyze all the elements in the sentence and know
how each word can influence another one.
Vauquois triangle
❏ In 1968, Bernard Vauquois examined the machine translation
approaches at his time to represent it in the form of a triangle. This
is known as the Vauquois triangle.
❏ The Vauquois triangle visualizes and describes the classical
approaches to machine translation, showing the evolution of those
approaches.
Vauquois triangle
Vauquois triangle
❏ These approaches are mainly, Direct approach, Transfer
approach, and Interlingua approach.
❏ The depth of analysis required for each approach increases as we
go from direct to interlingual, while the amount of transfer
knowledge needed decreases as we go from direct to interlingual
approach.
Direct translation
❏ The Direct Translation approach uses a bilingual dictionary to
translate the sentences word by word, where each source word is
mapped onto some target word.
❏ The translation process doesn’t use any intermediate structures,
other than some morphological analysis(the task of analyzing the
structure and parts of words) or lemmatization(the task of
converting a word to its basic dictionary form, for eg, translating to
translate).
Direct translation
❏ The source language is translated word by word using the bilingual
dictionary by mapping each word to its target language word, after
analyzing the structure and parts of the sentence. Then some
simple reordering processes are done on the resulting target
language sentence to make its structure correct.
Transfer-based translation
❏ The transfer-based translation approach first parses the source
language sentence to determine its structure, applies rules to
transfer the resulting structure to the target language parse
structure based on the knowledge of differences between the
languages, and then generates the target language sentence from
this transformed structure.
Transfer-based translation
❏ Based on the source and transfer languages, transfer approaches
can differ. For similar languages like French and Italian (almost 90%
lexically similar), only syntactic transfer is needed, where the
parsed source structure is transferred to the target parse structure
based on syntactic reordering.
❏ For complex translation processes like Korean to English, in
addition to syntactic transfer, we also need to do a semantic
transfer, where the ** source** structure needs to be transferred
according to their meaning and semantic roles in the target
sentence.
Interlingual translation
❏ The Interlingual approach first analyzes the source language
sentence, represents it as an Interlingua, and then generates the
target language sentence from this interlingua. Interlingua is a
language independent representation, that can be based on any
representation scheme.
❏ The availability of a large amount of data, and faster processing
systems changed the world of machine translation from the
classical approaches and the Vauquois era to faster and more
efficient results.
Basic Issues or Challenges Faced by Machine
Translation
1. Ambiguity

● Word Sense Ambiguity: Words can have multiple meanings

depending on context. For example, "bank" (financial institution vs.
riverbank).

● Structural Ambiguity: Sentence structure can lead to multiple

interpretations. E.g., "He saw the man with the telescope."
Basic Issues or Challenges Faced by Machine
Translation
2. Idioms and Expressions

● MT struggles with idiomatic expressions, metaphors, and slang

because they often don’t translate literally.

○ E.g., "Kick the bucket" shouldn't become "Patear el balde" in

Spanish.
Basic Issues or Challenges Faced by Machine
Translation
3. Cultural Context and Nuance

● Certain cultural references, humor, or tone don’t have direct

equivalents in other languages.

○ Example: Sarcasm or politeness levels (like in Japanese or

Korean) can be tough to replicate.
Basic Issues or Challenges Faced by Machine
Translation
4. Syntax and Grammar Differences

● Languages differ vastly in structure (e.g., Subject-Verb-Object in

English vs. Subject-Object-Verb in Japanese).

● Gender, tense, and case agreement can be inconsistent or

complex (especially in languages like Russian or German).
5. Low-Resource Languages

● For less commonly spoken or low-resource languages, there's

limited training data, leading to poor translation quality.

6. Named Entities and Proper Nouns

● MT can mistranslate names of people, places, organizations, or fail

to translate them when needed.

7. Domain Specificity

● General MT systems struggle with domain-specific texts (legal,

medical, technical, etc.) without specialized training.
8. Errors in Pronouns and Coherence

● Long sentences with multiple subjects or pronouns can confuse

MT systems, leading to mismatched or unclear references.

9. Bias and Offensive Language

● MT systems may inherit biases from training data or mistranslate

in ways that produce offensive or inappropriate outputs.

10. Real-Time and Speech Translation

● When translating spoken language in real time, accents, dialects,

and background noise introduce additional errors.
Statistical Machine Translation
For Statistical machine Translation and Word Alignment, please go
through following resource:
[Link]
Phrase based Translation
For Phrase-based Machine Translation, please go through following
resource:
[Link]