1. Getting the structure right for word alignment: LEAF Alexander Fraser and Daniel Marcu Presenter Qin Gao

2. Problem IBM Models have 1-N assumption Solutions A sophisticated generative story Generative Estimation of parameters Additional Solution Decompose the model components Semi-supervised training Result Significant Improvement on BLEU (AR-EN) Quick summary

3. The generative story Source word Head words Links to zero or more non-head words (same side) Non-head words Linked from one head word (same side) Deleted words No link in source side Target words Head words Links to zero or more non-head words (same side) Non-head words Linked from one head word (same side) Spurious words No link in target side

4. Minimal translational correspondence

6. The generative story ABC

7. 1a. Condition: Source word ABC

8. 1b. Determine source word class ABC

9. 2a. Condition on source classes C(A)C(B)C(C)

10. 2b. Determine links between head word and non-head words C(A)C(B)C(C)

11. 3a. Depends on the source head word ABC

12. 3b. Determine the target head word ABC X

13. 4a. Conditioned on source head word and cept size ABC X 2

14. 4b. Determine the target cept size ABC X 2 ?

15. 5a. Depend on the existing sentence length ABC X 2 ?

16. 5b. Determine the number of spurious target words ABC X 2 ??

17. 6a. Depend on the target word ABC X?? XYZXYZ

18. 6b. Determine the spurious word ABC X?Z XYZXYZ

19. 7a. Depends on target head word’s class and source word ABC C(X)?Z

20. 7b. Determine the non-head word it linked to ABC C(X)YZ

21. 8a. Depends on the classes of source/target head words C(A)BC C(X)YZ 123

22. 2 8b. Determine the position of target head word C(A)BC C(X) YZ 13

23. 2 8c. Depends on the target word class C(A)BC C(X) YZ 13

24. 32 8d. Determine the position of non-headwords C(A)BC C(X) Y Z 1

25. 132 9. Fill the vacant position uniformly C(A)BC C(X) Y Z

26. 132 (10) The real alignment C(A)BC C(X) Y Z

27. Unsupervised parameter estimation  Bootstrap using HMM alignments in two directions  Using the intersection to determine head words  Using 1-N alignment to determine target cepts  Using M-1 alignment to determine source cepts  Could be infeasible

28. Training: Similar to model 3/4/5  From some alignment (not sure how they get it), apply one of the seven operators to get new alignments  Move French non-head word to new head,  move English non-head word to new head,  swap heads of two French non-head words,  swap heads of two English non-head words,  swap English head word links of two French head words,  link English word to French word making new head words,  unlink English and French head words.  All the alignments that can be generated by one of the operators above, are called neighbors of the alignment

29. Training  If we have better alignment in the neighborhood, update the current alignment  Continue until no better alignment can be found  Collect count from the last neighborhood

30. Semi-supervised training  Decompose the components in the large formula treat them as features in log-linear model  And other features  Used EMD algorithm (EM-Discriminative) method

31. Experiment  First, a very weird operation, they fully link alignments from ALL systems and then compare the performance

32. Training/Test Set

33. Experiments  French/English: Phrase based  Arabic/English: Hierarchical (Chiang 2005)  Baseline: GIZA++ Model 4, Union  Baseline Discriminative: Only using Model 4 components as features

34. Conclusion(Mine)  The new structural features are useful in discriminative training  No evidence to support the generative model is superior over model 4

35. Unclear points  Are F scores “biased?”  No BLEU score given for LEAF unsupervised  They used features in addition to LEAF features, where is the contribution comes from?