Cohesive Constraints in a Beam Search Phrase-based Decoder

Interruption Checks (Cherry, 2008) la élection présidentielle commence demain des États Unis the presidential election begins tomorrow states united of the. 1.
772KB taille 3 téléchargements 437 vues
Cohesive Constraints in a Beam Search Phrase-based Decoder Nguyen Bach, Stephan Vogel Carnegie Mellon University

Colin Cherry Microsoft Research

1

Overview • Apply cohesive constraints during decoding process to consider the source dependency structures • Introduce extensions of the cohesive constraints. • Analyze the impact of cohesive constraints across language pairs with different reordering models • Applied to English-Spanish , English-Iraqi and ChineseEnglish translation tasks – Significant improvements on English-Spanish – Stable improvements on other pairs 2

Outline • Cohesive Decoding Approach • Experiments • Conclusions & Future Work

3

What is a cohesive decoding? begins election

1 the

presidential

Source dependency tree

2 tomorrow

3 of

states

the

united

English->French Source:

the presidential election of the united states begins tomorrow

la élection présidentielle

commence demain

des États Unis

4

What is a cohesive decoding? begins

the

3

election

1

presidential

Source dependency tree tomorrow

2 of

states

the

united

English->French Source:

the presidential election of the united states begins tomorrow

la élection présidentielle

des États Unis

commence demain

5

begins 1 the

2

election presidential

tomorrow 3 of

states the

Phrase-based decoder

united

the presidential election of the united states begins tomorrow la élection présidentielle

commence demain begins

1 the

3

election presidential

des États Unis

tomorrow 2 of

states the

united

Cohesive decoding

the presidential election of the united states begins tomorrow la élection présidentielle

des États Unis

commence demain

6

Interruption Checks (Cherry, 2008) begins election

tomorrow

presidential

the 1 2

3

states of

the

united

4 la

élection présidentielle commence demain des États Unis

7

Two Questions • How to determine the largest subtree that needs to be completed before the translation process can move elsewhere in the tree? – Interruption Check: use left and right most tokens of the previous translated source phrase and climb up the tree

• If a violation happens, how to constrain the decoder to penalize cohesion violated translation hypothesis? – Interruption Check : Binary event 8

Exhaustive Interruption Check • Interruption Check only penalizes the cohesion violation 1 time • Should penalties persist as long as violations remain unresolved? • Exhaustive Interruption Check keeps punishing a cohesion violation until it is fixed.

9

Exhaustive Interruption Check Exhaustive Interruption Check: YES

Interruption Check: NO begins

election presidential

the 1 2 la

tomorrow

3

5

states of

the

united

4

élection présidentielle commence demain des États Unis

10

Cohesion Violation Penalties • Interruption Check and Exhaustive Interruption Check: binary event • Are some violations worse than others? • Penalize a cohesion violation by the number of untranslated words under the largest subtree – Interruption Check -> Interruption Count – Exhaustive Interruption Check -> Exhaustive Interruption Count 11

Rich Interruption Constraints begins /VBZ

begins

OBJ

SBJ election

election /NN

tomorrow

NMOD the

presidential

states

the /DT

NMOD

tomorrow /NN NMOD

presidential /JJ PMOD

of

the

united

of /IN

the /DT

states /NNS NMOD

NMOD

united /VBN

• Penalize a cohesion violation by 4 constraints – – – –

Binary event: violation/not violate Interruption Count: untranslated word count Verb Count: untranslated verb count Noun Count: untranslated noun count 12

Comparison How to penalize a cohesion violation?

How to detect the largest subtree T(n)?

Binary

Number of untranslated words

Linguistics features

The previous phrase

Interruption Check

Interruption Count

Rich Interruption Constraints

All previous phrases

Exhaustive Interruption Check

Exhaustive Interruption Count

N/A

13

Outline • Cohesion Decoding Approach • Experiments • Conclusions & Future Work

14

English-Spanish; English-Iraqi English-Iraqi

33.4 33.2 33 32.8 32.6 32.4 32.2 32 31.8 31.6 31.4

BLEU

BLEU

English-Spanish

Europarl nctest2007

25 24.8 24.6 24.4 24.2 24 23.8 23.6 23.4 23.2 23

TransTac June08

Cohesive constraints obtained improvements over the standard phrase-based decoder. 15

How does the performance of the dependency parser affect cohesive constraints?

16

BLEU

The Role of Dependency Parser on English-Spanish 33.2 33 32.8 32.6 32.4 32.2 32 31.8 31.6 31.4 31.2

M1 M2

• Train 2 MALT dependency parser models: M1 with 10% of treebank and M2 with all treebank. • Performance on CoNLL-07 dependency test set – M1: 19.41% – M2: 86.21%

• Apply to MT – M2 is better than M1

17

• Are the improvements subsumed by a strong reordering model and system scale? • What if we translate from X->English?

18

GALE Chinese-English 27.2 27 26.8 26.6 26.4 26.2 26 25.8 25.6

GALE Dev07-WB

BLEU

BLEU

GALE Dev07-NW 25.5 25.4 25.3 25.2 25.1 25 24.9 24.8 24.7 24.6

Cohesive constraints obtained improvements even with large scale system and strong reordering models 19

Outline • Cohesion Decoding Approach • Experiments • Conclusions & Future Work

20

Conclusions & Future Work • Conclusions – Cohesive constraints are helpful – The effectiveness was shown when using with a strong reordering model – Obtained improvements with 3 language pairs and also covered a wide range of training corpus sizes, ranging from 500K up to 11M sentence pairs

• Future work – A source side dependency reordering model: Learning reordering events of the phrases based on source subtree movements – A hierarchical source side dependency reordering model: extend Galley&Manning (2008).

21

Questions

22