Comparing%20Corpus%20Co-Occurrence,%20Dictionary%20and%20Wikipedia%20Entries%20as%20Resources%20for%20Semantic%20Relatedness%20Information

1
Comparing Corpus Co-Occurrence, Dictionary and
Wikipedia Entries as Resources for Semantic
Relatedness Information

• Michael Roth

Sabine Schulte im Walde Universität
Stuttgart
2
Overview

• Motivation / Introduction
• Data-intensive lexical semantics
• Corpus-based descriptions
• Semantic Associations
• Our Work
• Evaluation of data-driven models
• Cross-comparison between resources
• Summary / Conclusions

3
Data-intensive lexical semantics

• Modelling word meaning
• Using meaning aspects
• Automatically obtainable
• Goal Determine (dis)similarity of words
• Applications
• Word sense discrimination
• Anaphora resolution
• ...

4
Corpus-based Descriptions

• Disadvantage Corpus co-occurrence does not cover
  all aspects of word meaning
• Especially world knowledge
• Our question Can we find complementing
  information in other resources?
• Dictionaries?
• Encyclopaedias?

5
Dictionary and Encyclopaedia

• Consider other resources
• Dictionaries contain detailed information
  about word senses
• Encyclopaedias written knowledge compendiums
• How to identify meaning aspects?
• In our work, we rely on semantic associations

6
Semantic Associations

• Definition
• We define semantic associations as concepts
  spontaneously called to mind by other concepts
  (stimuli)
• Assumption
• Evoked words reflect highly salient linguistic
  and conceptual features

7
Data Collection Verb Stimuli

• Associates to verb stimuli
• Web experiment
• 330 verb stimuli
• 30 seconds per verb

klagen complain, moan, sue klagen complain, moan, sue klagen complain, moan, sue
Gericht court 19
jammern moan 18
weinen cry 13
Anwalt lawyer 11
Richter judge 9
Klage complaint 7
Leid suffering 6
Trauer mourning 6
Klagemauer Wailing Wall 5
laut noisy 5
8
Data Collection Noun Stimuli

• Associates to noun stimuli
• Offline experiment
• 409 noun stimuli
• 3 associates per noun

Schloss castle, lock Schloss castle, lock Schloss castle, lock
Schlüssel key 51
Tür door 15
Prinzessin princess 8
Burg castle 8
sicher safe 7
Fahrrad bike 7
schließen close 7
Keller cellar 7
König king 7
Turm tower 6
9
Knowledge Resources

• Corpus data
• German newspaper corpus
• 200 mio. words
• Dictionary WDG
• (Wörterbuch der deutschen Gegenwartssprache)
• Freely available dictionary (130,000 entries)
• Average of 840 words/entry
• Encyclopedia Wikipedia
• Free online encyclopedia (650,000 articles)
• Average of 1,164 words/article

10
Analysis Vorgehensweise

• Corpus data
• Extract co-occurrence windows of stimuli
• Check windows for associations
• WDG / Wikipedia
• Download stimuli entries
• Check content for associations
• Missing entries
• WDG - 7/0
• Wikipedia - 2/54

11
Analysis Resource Coverage

• Noun associate (all)

• Verb associate (all)

POS Types Tokens
corpus 70 84
WDG 12 28
Wikipedia 26 46
POS Types Tokens
corpus 67 77
WDG 12 25
Wikipedia 6 10
1.2 2.3 1.8
1.2 2.0 1.7

• Resources differ in ...
• coverage per stimuli part-of-speech
• token/type ratio
• proportions per associates part-of-speech
  (next slide)

12
Analysis Resource Coverage (2)

• Proportions per associates part-of-speech
• Noun stimuli
• Corpus 88 V gt 84 N gt 83 Adj
• WDG 43 V gt 31 Adj gt 26 N
• Wikipedia 49 N gt 39 Adj gt 37 V
• Verb stimuli
• Corpus 91 Adv gt 79 V gt 77 Adj
  gt 76 N
• WDG 29 Adv gt 28 V gt 25 Ngt 24 Adj
• Wikipedia 12 N gt 9 Adj/Adv gt 6 V

13
Analysis Cross-Comparison

• Noun associate
• World knowledge?
• Only in WDG/Wiki carrot orange, cry tears,
  ...
• Only in Corpus igloo eskimo, teach school,
  ...

• Verb associate

Corpus Dic Wiki
Corpus - 55.0 46.0
WDG 0.8 - 5.7
Wiki 3.2 18.1 -
Corpus Dic Wiki
Corpus - 45.8 22.1
WDG 0.7 - 3.9
Wiki 0.5 3.6 -
14
Summary / Conclusions

• Analysis of associations across resources
• Results
• Different coverage per stimuli (noun vs. verb)
• Different (predominant) PoS in word descriptions
• Different strength of semantic relatedness
• Resources complement each other
• gt A combination of resources should be helpful
  for modelling word meaning and similarity

15
(No Transcript)
16
Questions?