Translation Theory

MACHINE
TRANSLATION
 CONCLUSION
The bilinguals Bilingualism
showed superior resulted in greater
performance on "mental flexibility"
certain cognitive and "abstract
tasks. thought".
TYPES OF
LEARNING SITUATIONS
Sequential
Learning

Simutaneous
Learning
Sequential Bilingualism
This occurs when a person
learns a second language
after they have already
begun to acquire their first
language.
 Example
Chinese child learning English
after acquiring Chinese
at home.
 Four typical stages
use home language even when others
don’t understand.
be silent, then use gesture.
understand parts of second language,
produce abbreviated utterances.
produce grammatical utterances in
appropriate situations
 Simultaneous Learning
Learning two languages at the same
time, usually from birth
Ex: A child who grows up with a
mother speaking English, a father
speaking Spanish
The Transfer Effect:
Language Relationships
Facilitation Across
Language
Different
Similarity Languages
 INTRODUCTION
The presentation on Machine
Translation (MT) begins by defining the
core concept: the use of computer
software to translate text or speech from
one language to another automatically.
MT is a sub-field of computational
linguistics that has evolved significantly
over the decades.
THE HISTORY OF MT
Before the 1940s: Some pioneering studies explored translation
automation.
1940s – mid-1960s: With the advent of the first computers,
several teams built operational MT systems. Approaches were
sometimes naïve but laid the groundwork for later
developments.
1990s: The emergence of Statistical Machine Translation (SMT)
based on large bilingual corpora, pioneered by IBM researchers
in the late 1980s. SMT became the basis of major systems like
Google Translate and Bing Translator.
Mid-2010s onward: Neural Machine Translation (NMT) using
deep learning revolutionized the field. Rising demand for online
translation brought MT back to the forefront of computational
linguistics.
 EARLY DEVELOPMENTS IN MT
(STATISTICAL MODELS)
Before SMT: Rule-based MT (RBMT) – handcrafted grammar
rules, limited scalability.

Emergence of SMT (late 1980s–1990s):

Based on probability theory and bilingual corpora.
Uses Bayes’ Theorem
 The Rise
The Rise of
of Neural
Neural Machine
Machine
Translation (NMT)
Translation (NMT)
Limitations of SMT: Improvements:
Breaking sentences into small More fluent translations with
chunks loses long-distance fewer grammatical errors.
semantic relationships. Better semantic preservation in
Difficulty handling rare words long sentences.
and low-resource languages. Stronger ability to capture
Often produces translations that context and nuance compared
sound unnatural or lack fluency to SMT.
 Case Study – Manning & Schütze’s
Foundational Work (1999)

Foundations of Statistical Natural

Language Processing (Manning &
Schütze, 1999) – While not solely
about MT, it is a foundational resource
for statistical NLP, with SMT as a key
component.
Neural vs. Statistical Machine
Translation
- Overview of statistical machine translation
(SMT)
- Neural networks in translation (Neubig’s
contributions)
- Case studies comparing SMT and NMT
 Overview of SMT

Statistical Machine Translation

(SMT) is an approach where The fundamental SMT formula,
translations are generated based based on Bayes' theorem
on statistical models derived from
bilingual text corpora

Translation Language
Model Model
That means
Word-based models (e.g., IBM Models
1–5)
Phrase-based models (more flexible,
captures short sequences instead of
single words)
Syntax-based models (incorporates
grammatical structures)
 Neural Networks in Translation
(Neubig’s Contributions)
a paradigm shift by using deep learning to directly model the
translation process as a single, end-to-end neural network.

Encoder–Decoder Architecture: Attention Mechanism:

Source sentences are encoded into Improves over fixed-length
continuous vector representations; encoding by letting the model
decoders generate target focus on relevant source words
sentences from these vectors. during translation.
Case studies
1 Translation Fluency
comparing SMT
and NMT 2 Accuracy

3 Data Requirements

4 Generalization
 SMT NMT

Translation Fluency

Accuracy

Data Requirements

Generalization
Source sentence: “The weather is nice today, let’s go for a walk.”
SMT Output: “Thời tiết thì tốt hôm nay, hãy đi cho một đi bộ.”
→ Rigid, word-for-word, unnatural Vietnamese.
NMT Output: “Hôm nay thời tiết đẹp, mình đi dạo nhé.”
→ Fluent, natural, and feels human-like.
Chapter 3 Evaluation and Challenges of MT

1/ We need fast, consistent ways to check MT quality.

2/ Human evaluation is accurate but slow & costly.

3/ Automatic metrics help developers improve systems daily.

 Metrics for evaluating
MT

BLEU METEOR
 BLEU
(Bilingual Evaluation Understudy)
1/ Measures: closeness to human translations.
2/ Method: Modified n-gram precision × Brevity Penalty.
3/ Strength: Simple, language-independent.
4/ Weakness: Focuses on surface match.
 BLEU
Examples
Ref: “It is a guide to action that ensures the military will heed Party
commands.”
Cand 1: “It is a guide to action which ensures that the military obeys
the party.” ✅ (high BLEU)
Cand 2: “It is to insure the troops forever hearing the activity
guidebook that party direct.” ❌ (low BLEU)
 BLEU
Examples
Ref: “It is a guide to action that ensures the military will heed Party
commands.”
Cand 1: “It is a guide to action which ensures that the military obeys
the party.” ✅ (high BLEU)
Cand 2: “It is to insure the troops forever hearing the activity
guidebook that party direct.” ❌ (low BLEU)
 METEOR
(Metric for Evaluation of Translation with Explicit Ordering)

1/ Measures: word matching with semantic awareness.

2/ Matches: exact, stems, synonyms, paraphrases.
3/ Uses: precision + recall + penalty for disordered matches.
4/ Strength: Captures meaning better than BLEU.
5/ Weakness: Slower, needs language tools.
 METEOR
(Examples)

Ref: “He purchased a car.”

Cand: “He bought an automobile.”
BLEU: low score (different words)
METEOR: high score (matches “purchased” ↔ “bought”,
“car” ↔ “automobile”
 II . Challenges in MT

Cultural
Idiom Context
nuances
 Cultural nuances
Example: Vietnamese “ăn Tết”
→ Correct: celebrate Lunar New Year / Wrong literal: eat Tet
BLEU METEOR

- Scores low for “eat Tet” because - Scores low for “eat Tet”
n-grams differ from reference.
because no synonym/semantic
- Cannot recognize cultural
adaptation beyond exact word
match with “celebrate Lunar
overlap. New Year”.
 Idiom
Example: Kick the bucket
→ Meaning: die / Literal: kick a bucket
BLEU METEOR

- Literal “kick a bucket” may - Literal “kick a bucket” scores low if

reference is “die” because no
still score high if reference
synonym link.
also contains “kick” and
- If synonym for “die” is in database,
“bucket”, even if meaning is METEOR can reward correct figurative
wrong. meaning.
 Context
Example: “He sat on the bank of the river”
→ Meaning: river shore / Wrong: financial bank
BLEU METEOR

- Better than BLEU if synonyms

- Wrong “financial bank”
like “shore” and “riverbank” are
translation scores low recognized.
 BLEU METEOR
Measures surface Adds synonym, stemming,

word overlap and paraphrase matching

→ better for meaning, but
→ quick but blind to
still not perfect for culture,
meaning errors.
idioms, or deep context.
 Chapter 4 Ethics and Future Trends in
Machine Translation

Ethical concerns in
MT The future of MT

Impact of MT on the translation profession

 Ethical concerns in MT
Popovic & Seligman's perspectives

1/ Bias & Translation Shifts

MT may preserve or amplify unnatural translation shifts
that harm target language quality.
Reducing all shifts to match evaluation metrics can make
translations less natural.
 Ethical concerns in MT
Popovic & Seligman's perspectives

2/ Transparency
MT systems work like “black boxes” → hard to explain why
errors occur.
3/ Evaluation Reliability
Using only one human reference in evaluation can skew
results due to translator-specific style.
 The future of MT

1/ Human-in-the-loop
Example: In legal translation, MT drafts a contract → lawyer-
translator checks terminology to ensure it matches jurisdiction
requirements.
2/ Interactive MT
Example: Translating a live news broadcast — MT suggests terms
instantly, translator edits for accuracy before airing.
 The future of MT
3/ Domain Adaptation
Example: MT adapted for finance → understands “bond” in
context as a debt instrument, not a glue or attachment.
4/ Post-editing as a profession
Example: E-commerce platforms use MT for product listings →
post-editors refine descriptions to avoid mistranslation that could
mislead buyers.
 Impact of MT on the
translation profession
Shift in skills
Job market changes
Increased productivity
Pressure on time & cost
..........
1/ The Growing Role of MT in Translation
MT is now a core tool in many translation workflows, from daily
communication to large-scale localization projects.
Advances in neural MT have significantly improved fluency and
accuracy.
MT enables faster turnaround and higher content volume,
especially for global businesses.
2/ Future Prospects & The Need for Human Involvement
MT will continue to evolve with domain adaptation, real-time
feedback, and semantic-aware evaluation.
The most effective model: Human + Machine collaboration — MT
handles speed and scale, humans ensure accuracy, nuance, and
ethical integrity.
 Discussion
1. What are contributions of MT?
2. Identify main features of NMT?
CONTRIBUTIONS TO MT

Detailed explanation of probabilistic principles in

language processing.
Development of bilingual translation models and
language models.
Techniques for word alignment – the backbone of SMT.
Provided formulas, examples, and SMT error analyses,
laying the groundwork for later research (e.g., Neubig,
2017 on NMT).
@reallygreatsite
Key Features:
End-to-end modeling: The entire translation process is
learned directly from bilingual data without breaking it into
separate modules.
Encoder–Decoder architecture:
Encoder converts the source sentence into a semantic
vector.
Decoder generates the target sentence from this
vector.
Attention mechanism (Bahdanau et al., 2015): Allows the
model to “focus” on relevant words/phrases during
translation.