Presentation is loading. Please wait.

Presentation is loading. Please wait.

MIXED SYNTAX TRANSLATION Hieu Hoang MT Marathon 2011 Trento.

Similar presentations


Presentation on theme: "MIXED SYNTAX TRANSLATION Hieu Hoang MT Marathon 2011 Trento."— Presentation transcript:

1 MIXED SYNTAX TRANSLATION Hieu Hoang MT Marathon 2011 Trento

2 Contents What is a syntactic model? Whats wrong with Syntax? Which syntax model to use? Why use syntactic models? Mixed-Syntax Model – Extraction – Decoding – Results Future Work

3 What is a syntactic model? Hierarchical Phrase-Based Model – String-to-string – Non-terminals are unlabelled Tree-to-string Model – Source non-terminals are labelled match input parse tree X habe X 1 gegessen # have eaten X 1 S habe NP 1 gegessen # have eaten NP 1

4 What is a syntactic model? Hierarchical Phrase-Based Model – String-to-string – Non-terminals are unlabelled Tree-to-string Model – Source non-terminals are labelled match input parse tree X habe X 1 gegessen # have eaten X 1 S habe NP 1 gegessen # have eaten NP 1

5 What is a syntactic model? Hierarchical Phrase-Based Model – String-to-string – Non-terminals are unlabelled Tree-to-string Model – Source non-terminals are labelled match input parse tree X habe X 1 gegessen # have eaten X 1 S habe NP 1 gegessen # have eaten NP 1

6 What is a syntactic model? Hierarchical Phrase-Based Model – String-to-string – Non-terminals are unlabelled Tree-to-string Model – Source non-terminals are labelled match input parse tree X habe X 1 gegessen # have eaten X 1 S habe NP 1 gegessen # have eaten NP 1

7 What is a syntactic model? Hierarchical Phrase-Based Model – String-to-string – Non-terminals are unlabelled Tree-to-string Model – Source non-terminals are labelled match input parse tree X habe X 1 gegessen # have eaten X 1 S habe NP 1 gegessen # have eaten NP 1

8 Contents What is a syntactic model? Whats Wrong with Syntax? Which syntax model to use? Why use syntactic models? Mixed-Syntax Model – Extraction – Decoding – Results Future Work

9 Whats Wrong with Syntax? (Ambati and Lavie, 2009) Evaluation of French-English MT System BLEUMETEOR Tree-to-string Tree-to-tree Moses (phrase-based)

10 Hierarchical Model according to János Veres, this would be in the first quarter of 2008 possible.

11 Hierarchical Model according to János Veres, this would be in the first quarter of 2008 possible.

12 Hierarchical Model according to János Veres, this would be in the first quarter of 2008 possible.

13 Tree-to-String Model lautJánosVereswärediesimerstenQuartal2008möglich. PN PP S TOP APPRNE VAFINPDSAPPRARTADJANNCARDADJA PUNC

14 Tree-to-String Model PN PP S TOP APPRNE VAFINPDSAPPRARTADJANNCARDADJA PUNC according toJánosVereswould bethisinthe firstquarter of2008possible.

15 Tree-to-String Model PN PP S TOP APPRNE VAFINPDSAPPRARTADJANNCARDADJA PUNC according toJánosVereswould bethisinthe firstquarter of2008possible. [X,1] [X,2]

16 Contents Whats Wrong with Syntax? Which syntax model to use? Why use syntactic models? Mixed-Syntax Model – Extraction – Decoding – Results Future Work

17 Other Syntactic Models Syntax-Augmented MT (SAMT) – Not constrained only to parse tree – (Zollmann and Venugopal, 2006) Binarization – Restructure and relabel parse parse tree – (Wang et al, 2010) Forest-based translation – Recover from parse errors – (Mi et al, 2008) Soft constraint – Reward/Penalize derivations which follows parse structure – (Chiang 2010)

18 Other Syntactic Models Syntax-Augmented MT (SAMT) – Not constrained only to parse tree – (Zollmann and Venugopal, 2006) Binarization – Restructure and relabel parse parse tree – (Wang et al, 2010) Forest-based translation – Recover from parse errors – (Mi et al, 2008) Soft constraint – Reward/Penalize derivations which follows parse structure – (Chiang 2010)

19 Other Syntactic Models Syntax-Augmented MT (SAMT) – Not constrained only to parse tree – (Zollmann and Venugopal, 2006) Binarization – Restructure and relabel parse parse tree – (Wang et al, 2010) Forest-based translation – Recover from parse errors – (Mi et al, 2008) Soft constraint – Reward/Penalize derivations which follows parse structure – (Chiang 2010)

20 Other Syntactic Models Syntax-Augmented MT (SAMT) – Not constrained only to parse tree – (Zollmann and Venugopal, 2006) Binarization – Restructure and relabel parse parse tree – (Wang et al, 2010) Forest-based translation – Recover from parse errors – (Mi et al, 2008) Soft constraint – Reward/Penalize derivations which follows parse structure – (Chiang 2010)

21 Other Syntactic Models Syntax-Augmented MT (SAMT) – Not constrained only to parse tree – (Zollmann and Venugopal, 2006) Binarization – Restructure and relabel parse parse tree – (Wang et al, 2010) Forest-based translation – Recover from parse errors – (Mi et al, 2008) Soft constraint – Reward/Penalize derivations which follows parse structure – (Chiang 2010)

22 Contents Whats Wrong with Syntax? Which syntax model to use? Why use syntactic models? Mixed-Syntax Model – Extraction – Decoding – Results Future Work

23 Why Use Syntactic Models? Decrease decoding time – Derivation constrained by source parse tree

24 Why Use Syntactic Models? Decrease decoding time – Derivation constrained by source parse tree Long-range reordering during decoding – rules covering more words than max-span limit

25 Why Use Syntactic Models? Decrease decoding time – Derivation constrained by source parse tree Long-range reordering during decoding – rules covering more words than max-span limit Other rule-forms – 3+ non-terminals – consecutive non-terminals – non-lexicalized rules

26 Why Use Syntactic Models? Decrease decoding time – Derivation constrained by source parse tree Long-range reordering during decoding – rules covering more words than max-span limit Other rule-forms – 3+ non-terminals – consecutive non-terminals – non-lexicalized rules X S 1 O 2 V 3 # S 1 V 3 O 2 X PRO 1 PRO 2 aime bien # PRO 1 like PRO 2

27 Contents Whats Wrong with Syntax? Which syntax model to use? Why use syntactic models? Mixed-Syntax Model – Extraction – Decoding – Results Future Work

28 Other Syntactic Models Syntax-Augmented MT (SAMT) – Not constrained only to parse tree – (Zollmann and Venugopal, 2006) Binarization – Restructure and relabel parse parse tree – (Wang et al, 2010) Forest-based translation – Recover from parse errors – (Mi et al, 2008) Soft constraint – Reward/Penalize derivations which follows parse structure – (Chiang 2010)

29 Other Syntactic Models Syntax-Augmented MT (SAMT) – Not constrained only to parse tree – (Zollmann and Venugopal, 2006) Binarization – Restructure and relabel parse parse tree – (Wang et al, 2010) Forest-based translation – Recover from parse errors – (Mi et al, 2008) Soft constraint – Reward/Penalize derivations which follows parse structure – (Chiang 2010) Ignore Syntax (occasionally)

30 Mixed-Syntax Model Tree-to-string model – input is a parse tree Roles of non-terminals – Constrain derivation to parse constituents – State information Consistent node label on target derivation hypotheses with different head NT cannot be recombined

31 Mixed-Syntax Model Tree-to-string model – input is a parse tree Roles of non-terminals – Constrain derivation to parse constituents Can sometime have no constraints – State information Consistent node label on target derivation hypotheses with different head NT cannot be recombined always X

32 Mixed-Syntax Model Naïve syntax model Mixed-Syntax Model Example Translation Rules VP VVFIN 1 zu VVINF 2 # to VVFIN 2 VVINF 1 VP X 1 zu VVINF 2 # X to X 2 X 1

33 Mixed-Syntax Model Naïve syntax model Mixed-Syntax Model Example Translation Rules VP VVFIN 1 zu VVINF 2 # to VVFIN 2 VVINF 1 VP X 1 zu VVINF 2 # X to X 2 X 1

34 Mixed-Syntax Model Naïve syntax model Mixed-Syntax Model Example Translation Rules VP VVFIN 1 zu VVINF 2 # to VVFIN 2 VVINF 1 VP X 1 zu VVINF 2 # X to X 2 X 1

35 Contents Whats Wrong with Syntax? Which syntax model to use? Why use syntactic models? Mixed-Syntax Model – Extraction – Decoding – Results Future Work

36 Extraction Allow rules – Max 3 non-terminals – Adjacent non-terminals At least 1 NT must be syntactic – Non-lexicalized rules

37 Example Rules Extracted RuleFactional Countp( t | s) Syntactic Rules VP NP 1 VVINF 2 # X X 2 X % Mixed Rules VP X 1 VZ 2 # X X 2 X % VP X 1 zu VVINF 2 # X to X 2 X % TOP NP 1 X 2 # X X 1 X %

38 Contents Whats Wrong with Syntax? Which syntax model to use? Why use syntactic models? Mixed-Syntax Model – Extraction – Decoding – Results Future Work

39 Synchronous CFG Input:

40 Synchronous CFG Input: Rules: S NP 1 VP 2 # NP 1 VP 2 NP je# I PRO lui# him VB vu# see VP ne PRO 1 VB 2 pas# did not VB 2 PRO 1

41 Synchronous CFG Input: Rules: S NP 1 VP 2 # NP 1 VP 2 NP je# I PRO lui# him VP ne PRO 1 VB 2 pas# did not VB 2 PRO 1 Derivation: VB vu# see

42 Synchronous CFG Input: Rules: S NP 1 VP 2 # NP 1 VP 2 NP je# I PRO lui# him VP ne PRO 1 VB 2 pas# did not VB 2 PRO 1 Derivation: VB vu# see

43 Synchronous CFG Input: Rules: S NP 1 VP 2 # NP 1 VP 2 NP je# I PRO lui# him VP ne PRO 1 VB 2 pas# did not VB 2 PRO 1 Derivation: VB vu# see

44 Synchronous CFG Input: Rules: S NP 1 VP 2 # NP 1 VP 2 NP je# I PRO lui# him VP ne PRO 1 VB 2 pas# did not VB 2 PRO 1 Derivation: VB vu# see

45 Synchronous CFG Input: Rules: S NP 1 VP 2 # NP 1 VP 2 NP je# I PRO lui# him VP ne PRO 1 VB 2 pas# did not VB 2 PRO 1 Derivation: VB vu# see

46 Synchronous CFG Input: Rules: S NP 1 VP 2 # NP 1 VP 2 NP je# I PRO lui# him VP ne PRO 1 VB 2 pas# did not VB 2 PRO 1 Derivation: VB vu# see

47 Mixed-Syntax Model Input:

48 Mixed-Syntax Model Input: Rules: S NP 1 VP 2 # X NP 1 VP 2 PRO je# X I PRO lui# X him VP ne X 1 pas# did not X 1 VB vu# X see X PRO 1 VB 2 # X X 2 X 1

49 Mixed-Syntax Model Input: Rules: S NP 1 VP 2 # X NP 1 VP 2 PRO je# X I PRO lui# X him VP ne X 1 pas# did not X 1 VB vu# X see X PRO 1 VB 2 # X X 2 X 1 Derivation:

50 Mixed-Syntax Model Input: Rules: S NP 1 VP 2 # X NP 1 VP 2 PRO je# X I PRO lui# X him VP ne X 1 pas# did not X 1 VB vu# X see X PRO 1 VB 2 # X X 2 X 1 Derivation:

51 Mixed-Syntax Model Input: Rules: S NP 1 VP 2 # X NP 1 VP 2 PRO je# X I PRO lui# X him VP ne X 1 pas# did not X 1 VB vu# X see X PRO 1 VB 2 # X X 2 X 1 Derivation:

52 Mixed-Syntax Model Input: Rules: S NP 1 VP 2 # X NP 1 VP 2 PRO je# X I PRO lui# X him VP ne X 1 pas# did not X 1 VB vu# X see X PRO 1 VB 2 # X X 2 X 1 Derivation:

53 Mixed-Syntax Model Input: Rules: S NP 1 VP 2 # X NP 1 VP 2 PRO je# X I PRO lui# X him VP ne X 1 pas# did not X 1 VB vu# X see X PRO 1 VB 2 # X X 2 X 1 Derivation:

54 Mixed-Syntax Model Input: Rules: S NP 1 VP 2 # X NP 1 VP 2 PRO je# X I PRO lui# X him VP ne X 1 pas# did not X 1 VB vu# X see X PRO 1 VB 2 # X X 2 X 1 Derivation:

55 Contents Whats Wrong with Syntax? Which syntax model to use? Why use syntactic models? Mixed-Syntax Model – Extraction – Decoding – Results Future Work

56 Experiment German-English Corpus GermanEnglish TrainSentences82,306 Words2,034,3731,965,325 TuneSentences2000 TestSentences1026 Trained:News Commentary 2009 Tuned:held out set Tested:nc test2007

57 Results Model# rules%BLEU Hierarchical61.2m15.9 Tree-to-String4.7m14.9 Mixed Syntax128.7m16.7 Using constituent parse German-English Model# rules%BLEU Hierarchical84.6m10.2 Mixed Syntax175.0m10.6 English-German

58 Example according to János Veres, this would be in the first quarter of 2008 possible. Hierarchical Model

59 Example Mixed-Syntax Model according to János Veres this would be possible in the first quarter of possible 2008quarter of [APPRART,1] [X,2] [CARD,3] thiswould be Veres János according to [X,1] [X,2] first [ADJA,1] [NN,2] in the [X,2] [PP,1] [PUNC,3] [PDS,2] [VAFIN,1] [NE,1] [NE,2] lautJánosVereswärediesimerstenQuartal2008möglich. APPRNE VAFINPDSAPPRARTADJANNCARDADJAPUNC PNPP S TOP

60 Example according to János Veres this would be possible in the first quarter of possible 2008quarter of [APPRART,1] [X,2] [CARD,3] thiswould be Veres János according to [X,1] [X,2] first [ADJA,1] [NN,2] in the [X,2] [PP,1] [PUNC,3] [PDS,2] [VAFIN,1] [NE,1] [NE,2] lautJánosVereswärediesimerstenQuartal2008möglich. APPRNE VAFINPDSAPPRARTADJANNCARDADJAPUNC PNPP S TOP Mixed Syntax

61 Example according to János Veres this would be possible in the first quarter of possible 2008quarter of [APPRART,1] [X,2] [CARD,3] thiswould be Veres János according to [X,1] [X,2] first [ADJA,1] [NN,2] in the [X,2] [PP,1] [PUNC,3] [PDS,2] [VAFIN,1] [NE,1] [NE,2] lautJánosVereswärediesimerstenQuartal2008möglich. APPRNE VAFINPDSAPPRARTADJANNCARDADJAPUNC PNPP S TOP Mixed Syntax

62 Example according to János Veres this would be possible in the first quarter of possible 2008quarter of [APPRART,1] [X,2] [CARD,3] thiswould be Veres János according to [X,1] [X,2] first [ADJA,1] [NN,2] in the [X,2] [PP,1] [PUNC,3] [PDS,2] [VAFIN,1] [NE,1] [NE,2] lautJánosVereswärediesimerstenQuartal2008möglich. APPRNE VAFINPDSAPPRARTADJANNCARDADJAPUNC PNPP S TOP Mixed Syntax

63 Example according to János Veres this would be possible in the first quarter of possible 2008quarter of [APPRART,1] [X,2] [CARD,3] thiswould be Veres János according to [X,1] [X,2] first [ADJA,1] [NN,2] in the [X,2] [PP,1] [PUNC,3] [PDS,2] [VAFIN,1] [NE,1] [NE,2] lautJánosVereswärediesimerstenQuartal2008möglich. APPRNE VAFINPDSAPPRARTADJANNCARDADJAPUNC PNPP S TOP Mixed Syntax

64 Chunk Tags Advantages of Shallow Tags – Dont need Treebank – More reliable Disadvantages – Not a tree structure We dont rely on tree structure

65 Results Shallow Tags German-English Model# rules%BLEU Hierarchical64.3m16.3 Mixed Syntax254.5m16.8

66 Larger Training Corpus German-English Corpus GermanEnglishCorpus TrainSentences1,446,224Europarl v5 Words37,420,87639,464,626 TuneSentences1910dev2006 Test (in-domain) (out-of- domain) Sentences nc test2007 v2 devtest2006

67 Larger Training Corpus German-English Model# rulesIn-domain (BLEU) Out-of-domain (BLEU) Hierarchical500m Mixed Syntax (original)2664m Mixed Syntax (new extraction) 1435m

68 Contents Whats Wrong with Syntax? Which syntax model to use? Why use syntactic models? Mixed-Syntax Model – Extraction – Decoding – Results Future Work

69 Create your own label Dumb labels ich bitte Sie, sich zu einer Schweigeminute zu erheben.

70 Create your own label Dumb labels ich bitte Sie, sich zu einer Schweigeminute zu erheben. S z e

71 Create your own labels Dumb labels ich bitte Sie, sich zu einer Schweigeminute zu erheben. S z e ModelIn-domain (BLEU) Out-of-domain (BLEU) Hierarchical Dumb Labels

72 Create your own labels Labels motivated by reordering Labelling patterns: 1.VMFIN...VVINF EOS 2.VVINF und... VVINF 3.VAFIN... (VVPP or VVINF) EOS 4., PRELS... VVINF EOS 5.EOS... zu VVINF Example: ich bitte Sie, sich zu einer Schweigeminute zu erheben. label 5 … werde ich dem Vorschlag von Herrn Evans folgen. label 3

73 Create your own labels Labels motivated by reordering ModelIn-domain (BLEU) Out-of-domain (BLEU) Hierarchical Dumb Labels Reordering Labels

74 Conclusion Mixed-Syntax Model – SCFG-based decoding – Hierarchical phrase-based v. tree-to-string – Generality v. specificity Syntax Models – Many variations – Wont automatically make MT better – Question which syntactic information? how do we use it? why use it?

75 END


Download ppt "MIXED SYNTAX TRANSLATION Hieu Hoang MT Marathon 2011 Trento."

Similar presentations


Ads by Google