Parsing Resources from Bilingual Texts

1
Parsing Resources from Bilingual Texts

• JHU 1998 Summer Workshop on Parsing Free Word
  Order Languages
• August 19, 1998
• Douglas Jones Cynthia Kuo

2
Czech Readers Digest Corpus

• Goal to use an aligned bilingual Czech-English
  corpus to build a monolingual Czech grammar.
• This summer survey and assessment
• Further work grammar building
• Text Czech Readers Digest Corpus
• About 1M words for each language
• Substantially pre-processed in Prague

3
Survey and Assessment

• Process both Czech and English texts with current
  parsers
• Large set 23K sentences
• look for (near)-isomorphic parses
• Small set 50 sentences
• inventory syntactically-motivated transformations
  

4
Sample Data
that 's strange , he thought . feeling the glass
, he discovered he could dampen the vibration but
could n't make it stop . i'm making it better
, '' les responded .

• to je zvlá¹tní , pomyslel si .
• sáhl na sklo a zjistil , ¾e chvìní mù¾e sice
  zmírnit , ale ¾e ho nezastaví úplnì .
• " vylep¹uju ho , " odpovìdìl les .

5
Closely Aligned Text

• Lengths of aligned sentences

6
Parsing Both Sides of Corpus

• Michael Collins Parser
• English (88 accuracy)
• trained on Penn Treebank
• converted output to dependency trees
• Czech (78 accuracy)
• trained on constituents derived from Prague
  Dependency Treebank
• yielded constituent and dependency trees

7
Exploration of Data Space
8
Data Issues

• Tagging
• Tokenizing
• Alignment
• Different treebank encodings
• Dependency versus constituency

9
Some tagset affinities

• Part of Czech English
• speech tag tag
• adjective A- J-
• adverb D- R-
• conjunction J- CC
• determiner --, (P-) D-
• noun N- N-
• preposition R- IN
• pronoun P- PRP
• verb V- V-

10
Transformable

• English Czech

11
Not (easily) transformable

• English Czech
• punctuation, paraphrasing, ...

12
Isomorphic Structures

• English Czech

13
Partially Matching Trees
14
Constituency Mismatches

• The English parses had richer constituent
  structure than the Czech parses.
• The Collins parser was trained on constituents
  derived automatically from the Czech dependency
  trees.

15
Further Research

• Grammar construction
• Fully process English text (get parse structures)
  
• Use lexicon (23K correspondences with
  probabilities, also string edit distance)
• Transformations for English trees to Czech trees
  (with lexical anchors)
• Applications for Machine Translation

16
Dependency Transformations

• Start by finding isomorphic matches
• Allow a transformation based on one mismatched
  position
• Collect new set of proposed matches as possible
  transformation
• Use greedy choice of transformations
• (We assume that the Czech-English correspondence
  is at best 70 reliable 88 English x 78
  Czech).

17
Dependency Transformations

• English Czech English Czech

Modal
Verb
Verb
18
Transformation Example
(a)
(a)
(b)

• (a) Initial match (2)
• (b) Match after transformation (5)

(a)
(b)
19
Brill Florian's Transformation Learning Tool

• English Czech
• feeling the glass feeling on glass
• feeling on glass
• Hypothesis learning Czech parse, given English
  parse,
• is easier than learning Czech parse from scratch

Take as much of the English parse as matches the
tags.
Apply transformation-driven learning, (with
hand-annotated examples.)
20
Vocabulary Correspondence

• About 12K words with likely translations (from
  word alignment experiment)
• total 1.000000e00
• absurdní 9
• 3.636364e-01 4 absurd
• 1.818182e-01 54 possible
• 1.818182e-01 2 preposterous
• 9.090909e-02 23044 ltEMPTY_WORD
• 9.090909e-02 7039 and
• Jan Hajic's students
• Cmejrek (1998) Master's Thesis on automatic
  extraction
• Curin (1988) Undergraduate Thesis on automatic
  extraction

21
Selecting Translation Guesses

• Minimal Edit Distance
• (Levinshtein Distance)
• 2.50
• 0.71
• 1.5
• 1.0

Czech English
Semantic Resources (WordNet) same synset (Wn
1.6)
22
Appendix

• Data Size
• Set Total Sample
• Size Size
• 010 1021 468
• 020 4006 2111
• 030 6115 1131
• 040 5317 1548