In this slides the basic concept of machine translation is described.MT challenges are represented and describes rule-based and statistical MT briefly. Some notes about evaluation is described too

1. Machine Translator

2. What is Machine Translator?
Automatic translation from one language to
another
Koehn: „Translating between languages is a task for
which even humans require special training.“

3. Why Even Humans Require Special Training
‫دادن‬ ‫نشان‬ ‫با‬«‫لولو‬»‫نمی‬ ‫قبله‬ ‫به‬ ‫رو‬ ‫ایران‬ ‫مردم‬ ،‫امنیت‬ ‫شورای‬ ‫ی‬‫ش‬‫وند‬
‫ترجمه‬‫نیوزویك‬:*‫گفته‬‫است‬‫كه‬‫اگر‬‫شورای‬‫امنیت‬‫مثل‬‫موجوداتی‬‫كه‬‫بچه‬‫ها‬‫را‬
‫می‬‫ترسانند‬‫ظاهر‬،‫شود‬‫مردم‬‫ایران‬‫به‬‫سوی‬‫قبله‬‫مسلمانان‬‫جهان‬‫دراز‬‫نمی‬‫كشند‬.
‫ترجمه‬‫نشریه‬‫اسپانیایی‬‫ال‬‫پائیس‬:*‫گفت‬‫كه‬‫اگر‬‫شورای‬‫امنیت‬‫چیز‬‫ترس‬‫ناكی‬‫را‬‫هم‬
‫به‬‫ایرانیان‬‫نشان‬،‫دهد‬‫باز‬‫هم‬‫مردم‬‫ایران‬‫به‬‫سوی‬‫عربستان‬‫سعودی‬‫نمی‬‫خوابن‬‫د‬.
‫ترجمه‬‫نشریه‬‫فرانسوی‬‫اومانیته‬:*‫گفت‬‫كه‬‫دراز‬‫كشیدن‬‫ایرانیان‬‫به‬‫سوی‬‫مركز‬
‫اعتقادات‬‫مسلمانان‬‫بستگی‬‫به‬‫این‬‫دارد‬‫كه‬‫آنها‬‫از‬‫موجودات‬‫افسانه‬‫ای‬‫بترسن‬،‫د‬‫این‬
‫یك‬‫داستان‬‫ایرانی‬‫است‬.

4. Other Similar Concept
•computer-aided translation
•machine-aided human translation (MAHT)

5. History
As early as the 17th century by philosophers René Descartes and
Gottfried Wilhelm Leibniz

6. MT in Computer

7. Applications
Dissemination
Publication in
other languages
Communication
Emails, chats
Assimilation
Understand the
Content

9. Challenges
Input
Typology
Lexical
Other

10. Input: Ambiguity
I saw a man with telescope

11. Input: Complexity
General relativity includes a dynamical spacetime so it is
difficult to see how to identify the conserved energy and
momentum Noether's theorem allows these quantities to be
determined from a Lagrangian with translation invariance but
general covariance makes translation invariance into
something of a gauge symmetry

12. Input: Wrong Sentence
I try for getting best grades but I did not can achive it

13. Typology: Morphology

14. Typology: Syntax
order of verbs (V), subjects (S) and objects (O)

15. Typology: Argument structure and linking

16. Typology: Pronouns omission

17. Lexical: Ambiguity

18. Lexical: Grammer

19. Lexical: Lexical gap

20. Lexical: Idiom

21. Other Challenges

22. Other Challenges

23. Human Translation Process
• Decoding the meaning of the source text
• Re-encoding this meaning in the target language.

24. Simplest Machine Translator
Apple ‫سیب‬

25. MT
Human
Translation
With Machine
Aid
Machine
Translation
With Human
Aid
Fully
Automated
Translation
Rule Based MT
Direct MT Transfer MT Interlingua
Knowledge
Based MT
Principle
Based MT
Empirical
Based MT
Statistical MT
Word Based
Translation
Phrase Based
Translation
Hierarchical
Phrase Based
Translation
Example
Based MT
Online
Interactive MT
Hybrid MT Neural MT

26. Rule-Base MT

27. Direct Translation
• dictionary has to cover all cross-lingual phenomena
• need to include contextual information in dictionary (long phrases)
• inflectional agreement, shifts in word order & structure
+direct translation systems include simplistic rules

28. Direct Translation Approach
• simplistic: only low-level pre/post-processing (tokenization, etc)
• advanced: handle some specific phenomena
identification & handling of syntactic ambiguity
morphological processing/synthesis
word re-ordering rules
rules for prepositions
handling of compounds and idioms, ...

29. Is Direct Translation Feasible?

30. Transfer Based Translation
Motivation:
• complete analysis of source language sentences
• handle lexical & structural ambiguity in one formalism

31. Transfer Based Needed Information/Tools
• source language parser (morpho-syntactic analysis)
• transfer engine (e.g. unification based grammar)
• target language generator

33. • Morphological analysis. Surface forms of the input text are classified as to part-
of-speech (e.g. noun, verb, etc.) and sub-category (number, gender, tense, etc.).
All of the possible "analyses" for each surface form are typically made output at
this stage, along with the lemma of the word.
• Lexical categorisation. In any given text some of the words may have more than
one meaning, causing ambiguity in analysis. Lexical categorisation looks at the
context of a word to try to determine the correct meaning in the context of the
input. This can involve part-of-speech tagging and word sense disambiguation.
• Lexical transfer. This is basically dictionary translation; the source language
lemma (perhaps with sense information) is looked up in a bilingual dictionary and
the translation is chosen.
• Structural transfer. While the previous stages deal with words, this stage deals
with larger constituents, for example phrases and chunks. Typical features of this
stage include concordance of gender and number, and re-ordering of words or
phrases.
• Morphological generation. From the output of the structural transfer stage, the
target language surface forms are generated.

34. Transfer Based: Syntactic Transfer

35. What are the problems?
• lots of grammar engineering (writing rules ...)
• language-pair specific rules
• exponential ambiguity
• variation & preference

36. Interlingua-based Translation

37. Persian
English

38. Persian
SpanishEnglish

39. Persian
Spanish
Japanese
English

40. interlingua
Persian
Spanish
New
English

41. Advantages & Disadvantages
• no language-pair specific transfer
• simple to add new languages (add new analysis/generation
component)
• need to design interlingua that covers all language phenomena
• need semantic representation (and that’s hard!)

43. Statistical MT

44. Statistical MT

45. Statistical MT

47. Statistical MT
(1) build a language model which allows us to estimate P(e)
(2) build a translation model which allows us to estimate P(f|e)
(3) search for e maximizing the product P(f|e).P(e)

48. Language Modeling

49. Which N-Gram?
• 1-Gram is not very realistic
• More realistic still is the trigram model
Problem
50,000 English word
2.5 billion possible bigrams
Many zero bigram in corpus but maybe needed in translations

50. linear interpolation

51. Translation Model
(i) a model of the sentence-aligned source–target training corpus
(ii) a method for computing the probability that S and T are
equivalent using that model

52. Translation Model Example

53. Example

57. Word Alignment

58. Simple Word Alignment

59. Expectation-Maximisation (EM) algorithm

60. Expectation-Maximisation (EM) algorithm

64. MT Evaluation
• How can we measure MT quality?
• How can we compare MT engines?
• How can we measure progress in MT development?

65. • Adequacy: Does the output convey the same meaning as the input sentence?
Is part of the message lost, added, or distorted?
• Fluency: Is the output good fluent English?
This involves both grammatical correctness and idiomatic word choices.

66. What do We Expect from MT?
• adequacy & informativeness (preserve meaning)
• fluency & grammaticality (translation needs to be natural)
• acceptance (for its task)

67. Task-specific evaluation
• browsing quality: Is the translation understandable in its
context?
• post-editing quality: How many edit operations are required
to turn it into a good translation?
• publishing quality: How many human interventions are
necessary to make the entire document ready for printing?

68. Evaluation is Difficult!
• I What is the best translation? (language variation!)
• I Subjective aspects (What is “fluent”? Clarity? Style?)
• I What is “grammatical”?
• I What is “adequate”? (Is it possible to be adequate?)

69. MT evaluation
Manual Evaluation
• ask actual users to rate translations
• statistics over user responses
• separate evaluations of adequacy &
fluency
• requires guidelines
• task-specific evaluation
Automatic Evaluation
• compare to reference translations
• approximations by measuring overlaps
• strong bias but useful for rapid
development

70. Fluency and Adequacy: Scales

72. Manual MT evaluation: What are the problems?
• need volunteers (every time we want to evaluate)
• expensive evaluation!
• subjective measures & disagreement between annotators

77. Automatic Evaluation: BLEU-score
• introduced in 2002 by Papineni et al
• desperately needed by rapid MT development
• quickly adapted by statistical MT community
• created a boom in MT research/experiments
• Many MT papers report only BLEU scores and don’t even look at the
translations

78. BLEU-score
the closer a machine translation is to a
professional human translation
the better it is

79. Definition
•Pn: for each pair of candidate and reference sentences.
• This score represents the proportion of n-word sequences in the
candidate translation which also occur in the reference translation.

81. • Koehn, Philipp. Statistical machine translation. Cambridge University Press, 2009.
• Arnold, D., et al. "Machine translation: An introductory guide. NCC Blackwell." (1994).
• https://www.slideshare.net/rushdishams/types-of-machine-translation
• https://en.wikipedia.org/wiki/Machine_translation
• Brown, Peter F., et al. "A statistical approach to machine translation." Computational
linguistics 16.2 (1990): 79-85.
• Hearne, Mary, and Andy Way. "Statistical machine translation: a guide for linguists and
translators." Language and Linguistics Compass 5.5 (2011): 205-226.

82. Question?