1. CS460/626 : Natural Language Processing/Speech, NLP and the Web (Lecture 18– Training and Decoding in SMT System) Kushal Ladha M.Tech Student CSE Dept., IIT Bombay 17 th Feb, 2011

2. Training Process It is a nine step process Prepare Data Run GIZA++ Align words Get lexical translation table Extract Phrases Score Phrases Build lexicalized reordering model Build generation models Create configuration file

3. Preparing Data (1/2) Sentence aligned data in two files. One file containing foreign sentences other containing English sentences Everything should be lowercase Sentence length less than 100 For factored model the training date should be Word0factor0|word0factor1|word0factor2 so on instead of the un-factored word0 word1 word 2 Cleaning the corpus drop empty lines, redundant spaces and eliminates sentence that violate 9-1 sentence ratio limit

4. Preparing Data (2/2) Input to GIZA++ is two vocabulary files Vocabulary files contain words, integer word identifiers and word count information 1UNK0 2and7 3,6 4irian4 Each sentence has 3 entries 1 47 74 65 56 70 5 81 7 84 5 75 2 24 69 35 7 55 44 3 86 5 46 2 36 15 16 22 4 71 49 39 11 60 48 57 17 68 53 55 29 13 44 3 25 61 2 36 70 5 GIZA++ requires each word to be placed into word classes Done by mkcls

5. Run GIZA++ Used to establish word alignments Word alignments are taken from the intersection of bidirectional runs of GIZA++ For each word in each sentence it marks the possible alignment points in both the direction

6. Align Words Heuristics GROW-DIAG-FINAL(e2f,f2e): neighboring = ((-1,0),(0,-1),(1,0),(0,1),(-1,-1),(-1,1),(1,-1),(1,1)) alignment = intersect(e2f,f2e); GROW-DIAG(); FINAL(e2f); FINAL(f2e); GROW-DIAG(): iterate until no new points added for english word e = 0... en for foreign word f = 0... fn if ( e aligned with f ) for each neighboring point ( e-new, f-new ): if (( e-new, f-new ) in union( e2f, f2e ) and ( e-new not aligned and f-new not aligned )) add alignment point ( e-new, f-new ) FINAL(a): for english word e-new = 0... en for foreign word f-new = 0... fn if ( ( ( e-new, f-new ) in alignment a) and ( e-new not aligned or f-new not aligned ) ) add alignment point ( e-new, f-new )

7. Get lexical Translation Table Given alignment, maximum likelihood lexical translation table can be estimated Two files in model/lex.0-0.f2e and model/lex.0-0.e2f contains the lexical translation probability Extract Phrases: model/extract contains all the phrases, with their translation in target language and alignment points An inverted file named model/extract.inv contains the inverse mapping

8. Score Phrases To calculate phrase translation probability sort the extract file, so that translation for a particular foreign word are next to each other in the file Calculate counts for each foreign word Do this for the inverted file to calculate for each English word Each phrase table entry consists of 5 probabilities Phrase translation probability Lexical weighting lex(f|e) Phrase translation probability Lexical weighting lex(e|f) Phrase penalty (exp(1)= 2.718 always )

9. Sample from Phrase-table b o ||| b aa ||| (0) (1) ||| (0) (1) ||| 1 0.666667 1 0.181818 2.718 b ||| b ||| (0) ||| (0) ||| 1 1 1 1 2.718 c o m p o ||| aa m p ||| (2) (0,1) (1) (0) (1) ||| (1,3) (1,2,4) (0) ||| 1 0.0486111 1 0.154959 2.718 c ||| p ||| (0) ||| (0) ||| 1 1 1 1 2.718 d w ||| d w ||| (0) (1) ||| (0) (1) ||| 1 0.75 1 1 2.718 d ||| d ||| (0) ||| (0) ||| 1 1 1 1 2.718 e b ||| ah b ||| (0) (1) ||| (0) (1) ||| 1 1 1 0.6 2.718 e l l ||| ah l ||| (0) (1) (1) ||| (0) (1,2) ||| 1 1 0.5 0.5 2.718 e l l ||| eh l ||| (0) (0) (1) ||| (0,1) (2) ||| 1 0.111111 0.5 0.111111 2.718 e l ||| eh ||| (0) (0) ||| (0,1) ||| 1 0.111111 1 0.133333 2.718 e ||| ah ||| (0) ||| (0) ||| 1 1 0.666667 0.6 2.718 h e ||| hh ah ||| (0) (1) ||| (0) (1) ||| 1 1 1 0.6 2.718 h ||| hh ||| (0) ||| (0) ||| 1 1 1 1 2.718 l e b ||| l ah b ||| (0) (1) (2) ||| (0) (1) (2) ||| 1 1 1 0.5 2.718 l e ||| l ah ||| (0) (1) ||| (0) (1) ||| 1 1 1 0.5 2.718 l l o ||| l ow ||| (0) (0) (1) ||| (0,1) (2) ||| 0.5 1 1 0.227273 2.718 l l ||| l ||| (0) (0) ||| (0,1) ||| 0.25 1 1 0.833333 2.718 l o ||| l ow ||| (0) (1) ||| (0) (1) ||| 0.5 1 1 0.227273 2.718 l ||| l ||| (0) ||| (0) ||| 0.75 1 1 0.833333 2.718 m ||| m ||| (0) ||| (0) ||| 1 0.5 1 1 2.718 n d ||| n d ||| (0) (1) ||| (0) (1) ||| 1 1 1 1 2.718 n e ||| eh n iy ||| (1) (2) ||| () (0) (1) ||| 1 1 0.5 0.3 2.718 n e ||| n iy ||| (0) (1) ||| (0) (1) ||| 1 1 0.5 0.3 2.718 n ||| eh n ||| (1) ||| () (0) ||| 1 1 0.25 1 2.718 o o m ||| uw m ||| (0) (0) (1) ||| (0,1) (2) ||| 1 0.5 1 0.181818 2.718 o o ||| uw ||| (0) (0) ||| (0,1) ||| 1 1 1 0.181818 2.718 o ||| aa ||| (0) ||| (0) ||| 1 0.666667 0.2 0.181818 2.718 o ||| ow eh ||| (0) ||| (0) () ||| 1 1 0.2 0.272727 2.718 o ||| ow ||| (0) ||| (0) ||| 1 1 0.6 0.272727 2.718 w o r ||| w er ||| (0) (1) (1) ||| (0) (1,2) ||| 1 0.1875 1 0.424242 2.718 w ||| w ||| (0) ||| (0) ||| 1 0.75 1 1 2.718

10. Build reordering, generation model Distance based reordering model is included Generation model is build in case of factored model It is done at the target side For eg: root/lemma at target side + suffix at target side  surface word at target side

11. Decoding It uses Beam search algorithm Translation options: given an input, a number of phrase translations are applied Translation options are collected before any decoding takes place Following information are stored along with TO First foreign word covered Last foreign word covered English phrase translation Phrase translation probability

12. Algorithm Start with null hypothesis For all the foreign words look for the translation options are expand it along with its probability The best probability hypothesis is selected and that foreign word is marked as translated We maintain back pointers of the hypothesis to read partial translations of the sentence The probability is nothing but cost of the new state i.e. cost of original state multiplied with the translation, distortion and language model cost of the added phrasal translation Final state is the state where all the foreign words are covered

13. Recombining Hypothesis Two hypothesis can be recombined if they agree in Foreign words covered so far The last two English words generated The end of the last foreign phrase covered We keep the cheaper hypothesis and discard the other one The pruning not only include the cost of each hypothesis so far but also the future estimate of the remaining sentence

14. Pruning Two types of pruning Threshold pruning a hypothesis which is less than α times the best hypothesis Histogram pruning Keeps certain number of hypothesis (e.g. n=100) This type of pruning is not risk free in comparison to recombining of hypothesis

15. Example Hindi Sentence is विनोद ने सचिन को छूरे से मारा | Required English sentence is Vinod Stabbed Sachin. At each level the best translation option is selected whose probability is highest of all. Hindi Sentence is विनोद ने सचिन को छूरे से मारा | Required English sentence is Vinod Stabbed Sachin. At each level the best translation option is selected whose probability is highest of all.