Analyzing Czech and English Parses Core NLP Technology Applicable to Multiple Languages

1
Analyzing Czech and English ParsesCore NLP
Technology Applicable to Multiple Languages

• Cynthia Kuo
• Stanford University
• 20 August 1998

2
Purpose

• Evaluation tool
• Compare a parser's output with the "truth"
• Use dependency structures in SGML format
• Output recall and precision
• Parsing by translation
• Explore the use of a bilingual corpus to modify
  English parsers for Czech

3
Truth Vs. Guess
ltfgtKanadskéltrgt1ltMDggt2 ltfgtfirmyltrgt2ltMDggt3 ltfgtmajíltr
gt3ltMDggt0 ltfgtzájemltrgt4ltMDggt0 ltfgtoltrgt5ltMDggt0 ltfgtspol
upráciltrgt6ltMDggt0 ltfgtastltrgt7ltMDggt0
ltfgtKanadskéltrgt1ltggt2 ltfgtfirmyltrgt2ltggt3 ltfgtmajíltrgt3ltg
gt0 ltfgtzájemltrgt4ltggt3 ltfgtoltrgt5ltggt4 ltfgtspolupráciltrgt6
ltggt5 ltdgtastltrgt7ltggt0
4
Evaluation

• Recall correct
  dependencies dependencies in truth
• 4/7 57.1
• Precision correct
  dependencies guesses by parser
• 4/7 57.1

• 7 dependencies in the truth
• 7 "guesses" by the parser
• 4 correct dependencies in parser's output

5
Parsing by Translation
with Doug Jones

• Ideally, the Czech and English sentences would
  look identical

6
Tag Affinities
Part of Czech English speech
tag tag adjective A- J- adverb D- R- conju
nction J- CC determiner --, (P-) D- noun N-
N- preposition R- IN pronoun P- PRP verb V-
V-

• Manually created
• Tags from
• Penn treebank
• Prague dependency treebank

7
Tag Affinities Unigram Frequencies
Content words POS Czech English adjective 655
643 adverb 833 920 noun 2112
2868 number 197 203 verb 2420 2442 total
number of words 9317 11214

• Set of aligned readers digest sentences
• Similar distribution of tags
• Slightly more English words than Czech words

8
More Unigram Frequencies

• Function words
• POS Czech English
• Conjunction 433 231
• Determiner -- 892
• Modal -- 173
• Particle 74 4
• Preposition 658 684
• Pronoun 1297 1262
• Unknown 531 22
• Total number
• Of words 9317 11214

• some words in English are not in Czech
• determiners
• modal verbs
• more unknown tags and particles in Czech

9
Transformations

• Eliminate punctuation
• Change word order / paraphrase
• Drop determiners
• Collapse modal verbs and infinitives
• Drop prepositions
• Drop subject

10
Transformations 2
English Czech
Word
Root

• Eliminate punctuation
• Change word order / paraphrase
• Drop determiners

Punc
Punc
1
1
2
3
2
3
11
Transformations 3
English Czech

• Collapse modal verbs and infinitives
• Drop prepositions
• Drop subject

Modal
Verb
Verb
12
Future Research Looking for Matches
13
Automating the Search?

• Compare domains?
• Grouping phrases
• Use domains containing descendents of uppermost
  ancestor
• Of , recording , music not that , recording
  , music
• Grouping around verbs
• Verbs play a central role
• Verbs usually appear in both sentences
• Matching groups by probability
• Identify important words in the phrases
• Make educated guesses about matching phrases
• Loosen affinities