The Interaction of Knowledge Sources in WSD

1
The Interaction of Knowledge Sources in WSD
Mark Stevenson and Yorik Wilks

• Davis Zhou
• College of Information Science Technology
  Drexel University

2
Outline

• Overview
• Word Sense in LDOCE
• Algorithm
• Part of Speech
• Dictionary Definition Overlap
• Selectional Preferences
• Subject Codes
• Feature Extractor
• Combining Disambiguation Modules.
• Evaluation

3
Overview

• WSD
• What kind of context information can be used?
• How to model these information?

4
Overview (Cont.)

• Architecture
• Preprocessing
• Feature Extractor
• One Filter Tagger Module remove some unlikely
  senses.
• Three Partial Tagger Modules evidence some very
  possible sense.
• One Combination Module

5
Overview (Cont.)

• Results
• Disambiguate all content words in text
• Use senses defined in LDOCE rather than WordNet
• Two corpus SEMCOR, SENSUS
• Fine-grained sense level 90 precision
  (monosemous words excluded)
• Coarse-grained homograph level 94 precision
  (monohomographic words excluded)

6
Sense in LDOCE

• Homograph
• Sets of senses with related meaning.
• Monohomographic (88) and polyhomographic (12)
• One homograph may contain more than one part of
  speech. (2 words in LDOCE contain a homograph
  with multiple POS)
• Sense
• 34 words are polysemous

7
Sense in LDOCE (Cont.)

• Example one homograph of Bank

8
Part of Speech

• Disambiguate Polyhomographic Words
• 88 words can be disambiguated to the homograph
  level since they dont have 2 homographs with the
  same POS.
• Additional 7 can be disambiguated if they are
  assigned a certain POS but not others.
• Experiment
• Wall Street Journal, 391 polyhomographic words
• 87.4 precision against baseline approachs (most
  frequently used sense) 78.

9
Part of Speech (Cont.)

• Further Analysis
• Full Disambiguation
• Partial Disambiguation
• No Disambiguation
• POS Error

10
Dictionary Definition Overlap

• Basic idea
• Using an overlap count of content words in
  dictionary definitions as a measure of semantic
  closeness. (Lesk, 1986)
• Optimization
• Computing overlap using the simulated annealing
  optimization algorithm (Cowage et al.,1992). 47
  sense level and 72 homograph level precisions.
• Improvement
• Normalize count by dividing the length of
  definition.
• Dont use pragmatic code.

11
Selectional Preferences

• Background
• LDOCE senses are marked with selectional
  restrictions expressed by 36 semantic codes.
• Bruce and Guthrie (1992) manually identified the
  relations.

12
Selectional Preference (Cont.)

• Mapping named entities onto LDOCE semantic codes.
• Identification of the sentence sites of
  relationship subject, direct and indirect
  object, noun modified by an adjective (shallow
  syntactic parser)
• An conservative algorithm return, for ambiguous
  noun, verb and adjective occurrences, the set of
  senses which satisfy the preferences imposed on
  them.

13
Selectional Preference (Cont.)
14
Subject Codes

• Basic idea
• Use training data to find weighted indicative
  words for each sense of a word. Then calculate
  the score of a window around the target word.
  Identify the sense with maximum score as the
  sense of the word.
• Yarowsky (1992)
• Collect contexts which are representative of the
  Roget category
• Groliers Encyclopedia 1991 as training corpus.

15
Subject Codes (Cont.)

• Concordances of 100 surrounding words for each
  occurrence of each member of the Rogets head
  (sense)
• Identify salient words in the collective context,
  and weight appropriately.

16
Subject Codes (Cont.)

• Use the resulting weights to predict the
  appropriate category for a word in novel context
• Papers Implementation
• Use British National Corpus
• Consider only words which appeared at least 10
  times in the training corpus.
• Use LDOCE pragmatic codes as category
• Return senses marked with the most likely
  pragmatic code.

17
Collocation Extractor

• A set of 10 collocates are extracted.
• First word to the left
• First word to the right
• Second word to the left
• Second word to the right
• First noun to the left
• First noun to the right
• First verb to the left
• First verb to the right
• First adjective to the left
• First adjective to the right

18
TiMBL

• Background
• Daelemans et al., 1999
• Used for various NLP tasks
• Classify
• Classifies new examples by comparing them against
  previously seen cases. The class of the most
  similar examples is assigned.

19
TiMBL (Cont.)

• Using training data to determine wi

20
TiMBL (Cont.)

• Feature Vector
• Head-word
• Homograph number
• Sense number
• Rank of sense
• Part of speech
• Surface form of the head-word
• Three partial taggers
• Ten collocates

21
TiMBL (Cont.)

• Examples

22
Evaluation

• Testing Material
• SEMCOR and SENSUS
• Converting SEMCORs WordNet sense to LDOCE sense.
  36869 words tagged with LDOCE senses.
• Occurrence of ambiguous words

23
Evaluation (Cont.)

• Average polysemy 14.62
• Testing technique 10-fold cross validation
• Evaluation metrics precision
• Baseline first sense algorithm

24
Evaluation (Cont.)

• Contribution of POS
• Accuracy at the sense level reduced to 87.87
• reduced to 93.36 at the homograph level when
  POS filter was removed.
• Interaction of Knowledge Sources.

25
Comments

• The context information the author used includes
  only previous word, next word, synonym,
  hypernym/hyponym. The limited use of context
  attributes to the low recall and high precision.
• Using clean topical signature as the context may
  improve the recall while keeping the high
  precision. The problem is it is highly expensive
  to build clean signature for each word.