Parser Adaptation and Projection with Quasi-Synchronous Grammar Features David A. Smith (UMass...

27
Parser Adaptation and Projection with Quasi-Synchronous Grammar Features David A. Smith (UMass Amherst) Jason Eisner (Johns Hopkins) 1

Transcript of Parser Adaptation and Projection with Quasi-Synchronous Grammar Features David A. Smith (UMass...

Page 1: Parser Adaptation and Projection with Quasi-Synchronous Grammar Features David A. Smith (UMass Amherst) Jason Eisner (Johns Hopkins) 1.

1

Parser Adaptation and Projection with Quasi-Synchronous Grammar Features

David A. Smith (UMass Amherst)Jason Eisner (Johns Hopkins)

Page 2: Parser Adaptation and Projection with Quasi-Synchronous Grammar Features David A. Smith (UMass Amherst) Jason Eisner (Johns Hopkins) 1.

2

This Talk in a Nutshell

in the beginning

im Anfang

Parser projection

German Spanish

Unsupervised

10 100 1000

Supervised

now or never

Parser adaptation

Learned by Quasi-Synchronous Grammar

Page 3: Parser Adaptation and Projection with Quasi-Synchronous Grammar Features David A. Smith (UMass Amherst) Jason Eisner (Johns Hopkins) 1.

3

Projecting Hidden Structure

Yarowsky & Ngai ‘01

Page 4: Parser Adaptation and Projection with Quasi-Synchronous Grammar Features David A. Smith (UMass Amherst) Jason Eisner (Johns Hopkins) 1.

Projection

• Train with bitext• Parse one side• Align words• Project dependencies• Many to one links?• Invalid trees?• Hwa et al.: fix-up rules• Ganchev et al.: trust only

some links

Im Anfang war das Wort

In the beginning was the word

Page 5: Parser Adaptation and Projection with Quasi-Synchronous Grammar Features David A. Smith (UMass Amherst) Jason Eisner (Johns Hopkins) 1.

Divergent Projection

Auf Fragediese bekommenichhabe leider Antwortkeine

I did not unfortunately receive an answer to this question

NULL

monotonicnull

head-swappingsiblings

Page 6: Parser Adaptation and Projection with Quasi-Synchronous Grammar Features David A. Smith (UMass Amherst) Jason Eisner (Johns Hopkins) 1.

Free Translation

Tschernobyl könnte dann etwas später an die Reihe kommen

Then we could deal with Chernobyl some time later

Bad dependencies

Parent-ancestors?

NULL

Page 7: Parser Adaptation and Projection with Quasi-Synchronous Grammar Features David A. Smith (UMass Amherst) Jason Eisner (Johns Hopkins) 1.

What’s Wrong with Projection?

• Hwa et al. Chinese data:– 38.1% F1 after projection– Only 26.3% with automatic

English parses– Cf. 35.9% for attach right!– 52.4% after fix-up rules

• Only 1-to-1 alignments:– 68% precision– 11% recall

Im Anfang war das Wort

In the beginning was the word

Page 8: Parser Adaptation and Projection with Quasi-Synchronous Grammar Features David A. Smith (UMass Amherst) Jason Eisner (Johns Hopkins) 1.

8

EXCURSUS: DOMAIN ADAPTATION

Page 9: Parser Adaptation and Projection with Quasi-Synchronous Grammar Features David A. Smith (UMass Amherst) Jason Eisner (Johns Hopkins) 1.

Projection

• Different languages• Similar meaning• Divergent syntaxIm Anfang war das Wort

In the beginning was the word

Page 10: Parser Adaptation and Projection with Quasi-Synchronous Grammar Features David A. Smith (UMass Amherst) Jason Eisner (Johns Hopkins) 1.

Adaptation

• Same sentence• Divergent syntax

In the beginning was the word

In the beginning was the word

Page 11: Parser Adaptation and Projection with Quasi-Synchronous Grammar Features David A. Smith (UMass Amherst) Jason Eisner (Johns Hopkins) 1.

11

A Lack of Coordination

now or never

Prague

now or never

Mel’čuk

now or never

CoNLL

now or never

MALT

Page 12: Parser Adaptation and Projection with Quasi-Synchronous Grammar Features David A. Smith (UMass Amherst) Jason Eisner (Johns Hopkins) 1.

12

Prepositions and Auxiliaries

in the end in the end in the end

I have decided I have decided

Page 13: Parser Adaptation and Projection with Quasi-Synchronous Grammar Features David A. Smith (UMass Amherst) Jason Eisner (Johns Hopkins) 1.

13

Adaptation Recipe

• Acquire (a few) trees in target domain• Run source-domain parser on training set• Train parser with features for:– Target tree alone– Source and target trees together

• Parse test set with:– Source-domain parser– Target-domain parser

Page 14: Parser Adaptation and Projection with Quasi-Synchronous Grammar Features David A. Smith (UMass Amherst) Jason Eisner (Johns Hopkins) 1.

14

Why?

• Why not just modify source treebank?• Source parser could be a black box– Or rule based

• Vastly shorter training times with a small target treebank– Linguists can quickly explore alternatives– Don’t need dozens of rules

• Other benefits of stacking• And sometimes, divergence is very large

Page 15: Parser Adaptation and Projection with Quasi-Synchronous Grammar Features David A. Smith (UMass Amherst) Jason Eisner (Johns Hopkins) 1.

15

MODEL STRUCTURE

Page 16: Parser Adaptation and Projection with Quasi-Synchronous Grammar Features David A. Smith (UMass Amherst) Jason Eisner (Johns Hopkins) 1.

16

What We’re Modeling

s(t, t ',a,w,w') = θ i f i(t,w)i

+ θ jg j (t, t ',a,w,w')j

in the beginning

im Anfang

t’w’

wt

a

p(t,a,w | t',w')

p(t | t',a,w,w')

This paper

Generative

Conditional

p(t, t ',a |w,w')Ongoing work

Page 17: Parser Adaptation and Projection with Quasi-Synchronous Grammar Features David A. Smith (UMass Amherst) Jason Eisner (Johns Hopkins) 1.

17

Stacking

Model 1Input

Model 2

Model 2 has features for when to trust Model 1

Page 18: Parser Adaptation and Projection with Quasi-Synchronous Grammar Features David A. Smith (UMass Amherst) Jason Eisner (Johns Hopkins) 1.

18

Quasi-Synchronous Grammar

• Generative or conditional monolingual model of target language or tree

• Condition target trees on source structure• Applications to– Alignment (D. Smith & Eisner ‘06)– Question Answering (Wang, N. Smith, Mitamura

‘07)– Paraphrase (Das & N. Smith ‘09)– Translation (Gimpel & N. Smith ‘09)

Page 19: Parser Adaptation and Projection with Quasi-Synchronous Grammar Features David A. Smith (UMass Amherst) Jason Eisner (Johns Hopkins) 1.

Dependency Relations

+ “none of the above”

Page 20: Parser Adaptation and Projection with Quasi-Synchronous Grammar Features David A. Smith (UMass Amherst) Jason Eisner (Johns Hopkins) 1.

21

EXPERIMENTS

Page 21: Parser Adaptation and Projection with Quasi-Synchronous Grammar Features David A. Smith (UMass Amherst) Jason Eisner (Johns Hopkins) 1.

22

Experimental Plan

• Proof of concept on English dependency-convention adaptation

• Unsupervised projection– No target trees– Generative target model + QG features

• Supervised projection– Small number of target trees– Conditional target model + QG features

Page 22: Parser Adaptation and Projection with Quasi-Synchronous Grammar Features David A. Smith (UMass Amherst) Jason Eisner (Johns Hopkins) 1.

23

Adaptation Results

Parsed source Gold source Parsed match Gold match0.6

0.65

0.7

0.75

0.8

0.85

0.9

0.95

1

010100

See paper for more resultsDifferent PTB dep. conversions

Page 23: Parser Adaptation and Projection with Quasi-Synchronous Grammar Features David A. Smith (UMass Amherst) Jason Eisner (Johns Hopkins) 1.

24

Unsupervised Projection

German Spanish0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

Mod. Prev.Mod. NextMonolingual EMHard proj.Hard proj. w/EMQG w/EM

Page 24: Parser Adaptation and Projection with Quasi-Synchronous Grammar Features David A. Smith (UMass Amherst) Jason Eisner (Johns Hopkins) 1.

25

Supervised Projection

10 100 10000.58

0.63

0.68

0.73

0.78

0.83

0.88

Target onlySource textGold alignmentsGold parsesGold parses, alignments

Page 25: Parser Adaptation and Projection with Quasi-Synchronous Grammar Features David A. Smith (UMass Amherst) Jason Eisner (Johns Hopkins) 1.

26

Conclusions

• Unified adaptation and projection• Conditional and generative training with

quasi-synchronous grammar features• Learned regular divergences• Ongoing work:– Joint, but not synchronous, inference– Better alignments– More adaptation problems

Page 26: Parser Adaptation and Projection with Quasi-Synchronous Grammar Features David A. Smith (UMass Amherst) Jason Eisner (Johns Hopkins) 1.

27

QUESTIONS?Thanks!

Page 27: Parser Adaptation and Projection with Quasi-Synchronous Grammar Features David A. Smith (UMass Amherst) Jason Eisner (Johns Hopkins) 1.

28

Really Different Domains

中国 在 基本 建设 方面 , 开始

In the area of infrastructure construction , China has begun…