Using Percolated Dependencies in PBSMT

1
Using Percolated Dependencies in PBSMT
CLUKI XII April 24, 2009

• Ankit K. Srivastava and Andy Way
• Dublin City University

2
About
3
Syntactic Parsing and Head Percolation
4
Parsing I Constituency Structure

• Vinken will join the board as a nonexecutive
  director Nov 29
• (ROOT
• (S
• (NP (NNP Vinken))
• (VP (MD will)
• (VP (VB join)
• (NP (DT the) (NN board))
• (PP (IN as)
• (NP (DT a) (JJ
  nonexecutive) (NN director)))
• (NP (NNP Nov) (CD 29))))))

5
Parsing II Dependency Structure

• Vinken will join the board as a nonexecutive
  director Nov 29
• HEAD DEPENDENT
• join Vinken
• join will
• board the
• join board
• join as
• director a
• director nonexecutive
• as director
• 29 Nov
• join 29

6
Parsing III Head Percolation

• It is straightforward to convert constituency
  tree to an unlabeled dependency tree (Gaifman
  1965)
• Use head percolation tables to identify head
  child in a constituency representation (Magerman
  1995)
• Dependency tree is obtained by recursively
  applying head child and non-head child heuristics
  (Xia Palmer 2001)
• (NP (DT the) (NN board))
• NP right NN/NNP/CD/JJ
• (NP-board (DT the) (NN board))
• the is dependent on board

7
Parsing IV Three Parses

• Constituency (phrase-structure) parses
  CONrequires CON parser
• Dependency (head-dependent) parses DEPrequires
  DEP parser
• Percolated (head-dependent) parses
  PERCrequires CON parser heuristics

8
Phrase-Based Statistical Machine Translation
9
PBSMT I Framework

• argmaxe p(ef) argmaxe p(fe) p(e)
• Decoder, Translation Model, Language Model
• PBSMT framework in Moses (Koehn et al., 2007)
• Phrase Table in Translation Model Align
  words extract phrases score phrases
• Different methods to extract phrases
• Moses phrase extraction as baseline system

10
PBSMT II Non-syntactic Phrase Extraction

• baseline Moses
• Get word alignments (src2tgt, tgt2src)
• Perform grow-diag-final heuristics (Koehn et al.,
  2003)
• Extract phrase pairs consistent with the word
  alignments
• String-based (non-syntactic) phrases STR

11
PBSMT III Syntactic Phrase Extraction

• Get word alignments (src2tgt, tgt2src)
• Parse src sentences
• Parse tgt sentences
• Use Tree Aligner to align subtree nodes (Zhechev
  2009)
• Extract surface-level chunks from parallel
  treebanks
• Previously, Tinsley et al., 2007 Hearne et al.,
  2008
• Syntactic phrases
• CON DEP PERC

12
System Design
13
System I Tools and Resources

• English-French parallel corpora
• Phrase Structure Parsers (En, Fr)
• Dependency Structure Parsers (En, Fr)
• Head Percolation tables (En, Fr)
• Statistical Tree Aligner
• Giza Word Aligner
• SRILM (Language Modeling) Toolkit
• Moses Decoder

14
System II Entries in Phrase tables Europarl
PERC is a unique knowledge source
but is it useful?
15
System III Combinations

• Concatenate phrase tables and re-estimate
  probabilities
• 15 different systems ?4Cr , 1r4
• STR CON DEP
  PERC

16
MT Systems and Evaluation
17
Numbers I Evaluation - JOC
18
Numbers II Evaluation - Europarl
19
Numbers III Uniquely best

• Evaluate MT systems STR, CON, DEP, PERC on a per
  sentence level. (Translation Error Rate)
• JOC (440 sentences)
• Europarl (2000 sentences)

20
Numbers IV Adding PERC Europarl
21
Analysis of Results
22
Analysis I STR

• Using Moses baseline phrases (STR) is essential
  for coverage. SIZE matters!
• However, adding any system to STR increases
  baseline score. Symbiotic!
• Hence, do not replace STR, but augment it.

23
Analysis II CON

• Seems to be the best combination with STR (SC
  seems to be the best performing system)
• Has most common chunks with PERC
• Does PERC harm a CON system needs more analysis

24
Analysis III DEP

• PERC is different from DEP chunks, despite being
  formally equivalent
• PERC can substitute DEP

25
Analysis IV PERC

• Is a unique knowledge source.
• Sometimes, it helps.
• Needs more work on finding connection with CON /
  DEP

26
Conclusion Future Work
27
Conclusion Future Work

• Extended Hearne et al., 2008 by- scaling up data
  size from 7.7K to 100K- introducing percolated
  dependencies in PBSMT
• Manual evaluation
• More analysis of results
• More combining strategies
• Seek to determine if each chunk type owns
  sentence types

28
Thanks

• ltasrivastava _at_ computing.dcu.iegt