1. Corpora and Translation Parallel corpora Statistical MT (not to mention: Corpus of translated text, for translation studies)

2. 2/38 Parallel corpora Corpora of texts and their translations Basic idea that such parallel corpora implicitly contain lots of information about translation equivalence Nowadays many such “bitexts” are available –bilingual countries have laws, parliamentary proceedings, and other documents –large multinational organizations (UN, EU [Europarl corpus], etc.) –multinational commercial organizations produce multilingual texts

3. 3/38 Bilingual concordance Source: TransSearch, Laboratoire de Recherche Appliquée en Linguistique Informatique, Université de Montréal http://www-rali.iro.umontreal.ca

4. 4/38 Parallel corpora Usually not corpora in the strict sense (planned, annotated, etc.) Usefulness may depend on –the quality of translation –the closeness of translation –whether we have a text and its translation, or a multilingually authored text –the language pair Parallel corpus needs to be aligned

5. 5/38 Alignment Means annotating the bilingual corpus to show explicitly the correspondences –at sentence level –at word and phrase level Main difficulty for sentence alignment is that translations do not always keep sentence boundaries, or even sentence order In addition, translation may be “localized” and therefore not especially faithful

6. 6/38 Sentence-level alignment If parallel corpus is quite a literal translation, this can be done using quite low-level information –sentence length –looking for anchors proper names, dates, figures eg in a parliamentary debate, speakers’ names

7. 7/38 Alignment tools

8. 8/38 Corpus-based MT Translation memory (tool for translators) –database of previous translations –find close matching examples to current translation unit –translator decides what to do with it

9. 9/38 Note that translator has to know/decide what bits of the target sentence to change

10. 10/38 Corpus-based MT Translation memory (tool for translators) –database of previous translations –find close matching examples to current translation unit –translator decides what to do with it Example-based translation –similar idea, but computer program tries to manipulate example(s) –may involve “learning” general rules from multiple examples

11. 11/38 Statistical MT Pioneered by IBM in early 1990s Spurred on by better success in speech recognition of statistical over linguistic rule- based approaches Idea that translation can be modelled as a statistical process Seems to work best in limited domain where given data is a good model of future translations

12. 12/38 Translation as a probabilistic problem For a given SL sentence S i, there are  number of “translations” T of varying probability Task is to find for S i the sentence T j for which the probability P(T j | S i ) is the highest

13. 13/38 Two models P(T j | S i ) is a function of two models: –The probabilities of the individual words that make up T j given the individual words in S i - the “translation model” –The probability that the individual words that make up T j are in the appropriate order – the “language model”

14. 14/38 Expressed in mathematical terms: Since S is a given, and constant, this can be simplified as Translation modelLanguage model

15. 15/38 So how do we translate? For a given input sentence S i we have to have a practical way to find the T j that maximizes the formula We have to start somewhere, so we start with the translation model: which words look most likely to help us? In a systematic way we can keep trying different combinations together with the language model until we stop getting improvements

16. 16/38 Input sentence Translation model Bag of possible words Most probable translation Seek improvement by trying other combinations Language model

17. 17/38 Where do the models come from? All the statistical parameters are pre-computed (“learned”), based on a parallel corpus Language model is probabilities of word sequences (n-grams) Translation model is derived from aligned parallel corpus This approach is attractive to some as an example of “machine learning” –The computer learns to translate (just) from seeing previous examples of translation

18. 18/38 The translation model Take sentence-aligned parallel corpus Extract entire vocabulary for both languages For every word-pair, calculate probability that they correspond – e.g. by comparing distributions

19. 19/38 Problem: fertility “fertility”: not all word correspondences are 1:1 –Some words have multiple possible translations, e.g. the  {le, la, l’, les} –Some words have no translation, e.g. in il se rase ‘he shaves’, se  –Some words are translated by several words, e.g. cheap  peu cher –Not always obvious how to align

20. 20/38 Problem: distortion Notice that corresponding words do not appear in the same order. The translation model includes probabilities for “distortion” –e.g. P(2|5): the P that w s in position 2 will produce a w t in position 5 –can be more complex: P(5|2,4,6): the P that w s in position 2 will produce a w t in position 5 when S has 4 words and T has 6.

21. 21/38 The language model Impractical to calculate probability of every word sequence: –Many will be very improbable … –Because they are ungrammatical –Or because they happen not to occur in the data Probabilities of sequences of n words (“n- grams”) more practical –Bigram model: where P(w i |w i–1 )  f(w i–1, w i )/f(w i )

22. 22/38 Sparse data Relying on n-grams with a large n risks 0- probabilities Bigrams are less risky but sometimes not discriminatory enough –e.g. I hire men who is good pilots 3- or 4-grams allow a nice compromise, and if a 3-gram is previously unseen, we can give it a score based on the component bigrams (“smoothing”)

23. 23/38 Put it all together and …? To build a statistical MT system we need: –Aligned bilingual corpus –“Training programs” which will extract from the corpora all the statistical data for the models –A “decoder” which takes a given input, and seeks the output that evaluates the magic argmax formula – based on a heuristic search algorithm Software for this purpose is freely available –http://www.statmt.org/moses/, http://www.isi.edu/licensed-sw/pharaoh/ Claim is that an MT system for a new language pair can be built in a matter of hours

24. 24/38 SMT latest developments Nevertheless, quality is limited SMT researchers quickly learned that this crude approach can get them so far (quite far actually), but that to go the extra distance you need linguistic knowledge (eg morphology, “phrases”, consitutents) Latest developments aim to incorporate this Big difference is that it too can be LEARNED (automatically) from corpora So SMT still contrasts with traditional RBMT where rules are “hand coded” by linguists

25. 25/38 Direct phrase alignment (Wang & Waible 1998, Och et al., 1999, Marcu & Wong 2002) Enhance word translation model by adding joint probabilities, i.e. probabilities for phrases Phrase probabilities compensate for missing lexical probabilities Easy to integrate probabilities from different sources/methods, allows for mutual compensation

26. 26/38 Word alignment induced model Koehn et al. 2003; example stolen from Knight & Koehn http://www.iccs.inf.ed.ac.uk/~pkoehn/publications/tutorial2003.pdf http://www.iccs.inf.ed.ac.uk/~pkoehn/publications/tutorial2003.pdf Maria did not slap the green witch Maria no daba una botefada a la bruja verda Start with all phrase pairs justified by the word alignment

27. 27/38 Word alignment induced model Koehn et al. 2003; example stolen from Knight & Koehn http://www.iccs.inf.ed.ac.uk/~pkoehn/publications/tutorial2003.pdf http://www.iccs.inf.ed.ac.uk/~pkoehn/publications/tutorial2003.pdf (Maria, Maria), (no, did not) (daba una botefada, slap), (a la, the), (verde, green), (bruja, witch)

28. 28/38 Word alignment induced model Koehn et al. 2003; example stolen from Knight & Koehn http://www.iccs.inf.ed.ac.uk/~pkoehn/publications/tutorial2003.pdf http://www.iccs.inf.ed.ac.uk/~pkoehn/publications/tutorial2003.pdf (Maria, Maria), (no, did not) (daba una botefada, slap), (a la, the), (verde, green) (bruja, witch), (Maria no, Maria did not), (no daba una botefada, did not slap), (daba una botefada a la, slap the), (bruja verde, green witch) etc.

29. 29/38 Word alignment induced model Koehn et al. 2003; example stolen from Knight & Koehn http://www.iccs.inf.ed.ac.uk/~pkoehn/publications/tutorial2003.pdf http://www.iccs.inf.ed.ac.uk/~pkoehn/publications/tutorial2003.pdf (Maria, Maria), (no, did not), (slap, daba una bofetada), (a la, the), (bruja, witch), (verde, green), (Maria no, Maria did not), (no daba una bofetada, did not slap), (daba una bofetada a la, slap the), (bruja verde, green witch), (Maria no daba una bofetada, Maria did not slap), (no daba una bofetada a la, did not slap the), (a la bruja verde, the green witch), (Maria no daba una bofetada a la, Maria did not slap the), (daba una bofetada a la bruja verde, slap the green witch), (no daba una bofetada a la bruja verde, did not slap the green witch), (Maria no daba una bofetada a la bruja verde, Maria did not slap the green witch)

30. 30/38 Alignment templates Och et al. 1999; further developed by Marcu and Wong 2002, Koehn and Knight 2003, Koehn et al. 2003) Problem of sparse data worse for phrases So use word classes instead of words –alignment templates instead of phrases –more reliable statistics for translation table –smaller translation table –more complex decoding Word classes are induced (by distributional statistics), so may not correspond to intuitive (linguistic) classes Takes context into account

31. 31/38 Problems with phrase-based models Still do not handle very well... –dependencies (especially long-distance) –distortion –discontinuities (e.g. bought = habe... gekauft ) More promising seems to be...

32. 32/38 Syntax-based SMT Better able to handle –Constituents –Function words –Grammatical context (e.g. case marking) Inversion Transduction Grammars Hierarchical transduction model Tree-to-string translation Tree-to-tree translation

33. 33/38 Inversion transduction grammars Wu and colleagues (1997 onwards) Grammar generates two trees in parallel and mappings between them Rules can specify order changes Restriction to binary rules limits complexity

34. 34/38 Inversion transduction grammars

35. 35/38 Inversion transduction grammars Grammar is trained on word-aligned bilingual corpus: Note that all the rules are learned automatically Translation uses a decoder which effectively works like traditional RBMT: –Parser uses source side of transduction rules to build a parse tree –Transduction rules are applied to transform the tree –The target text is generated by linearizing the tree

36. 36/38

37. 37/38

38. 38/38 Other approaches Other approaches use more and more “linguistic” information In each case automatically learned, especially from treebanks Traditional (“rule-based”) MT used (hand- written) grammars and lexicons State-of-the-art MT is moving back in this direction, except that linguistic rules are machine learned