Word Sense Disambiguation in Queries

1
Word Sense Disambiguation in Queries

• Shuang Liu, Clement Yu
• University of Illinois at Chicago
• Weiyi Meng
• Binghamton University
• CIKM 2005

2
Abstract

• This paper present a new approach to determine
  the senses of words in queries by using WordNet.
• Noun phrases in a query are determined first.
• Pieces of information associated with query words
  are compared to assign senses to these words.
• A guess and a web search are applied if
  necessary.
• Experimental results show that this approach has
  100 applicability and 90 accuracy for WSD.

3
Introduction

• WSD
• Several approaches such as machine learning and
  dictionary based ones are developed.
• The best reported result for WSD is 71.2.
  Mihalcea, 2002
• WSD is important for IR
• Adding appropriate synonyms ad hyponyms to a
  query can improve retrieval effectiveness.
• Adding synonyms of incorrect senses of a query
  term will deteriorate IR performance.

4
A Brief Introduction to WordNet

• A term w has possibly many sets (senses) of
  synonyms
• Each set S(w)i is called a synset in WordNet
• Each synset is associated with a definition D(w)i
• Each synset may have a number of hyponym synsets
• Each hyponym synset H(w)ij have the same but
  narrower meaning
• D(H(w)ij) represents the definition of H(w)ij
• Each synset may belong to a domain (category)
  Dom(w)i
• Ex hit1 belongs to baseball1

5
The WSD Algorithm

• Noun phrases (w, w) in the query are detected.
• In WordNet, the pieces of information of the
  words which form a phrase are compared to assign
  senses to these words.
• Synonyms, hyponyms, definitions of synonyms and
  hyponyms, and domains.
• If the sense of a query word has not been
  identified, information of other query words are
  used by through the same process as II.
• If the sense of a query word has not been
  identified, a guess is applied if it has at least
  50 chance of being correct.
• A Web search is applied if there are still some
  words whose senses have not been determined.

6
Case Analysis of Comparing WordNet Information

• Case 1
• w and w have a common synonym w.
• ? The senses of w and w are the synset
  containing w
• Full matching both w and w have the same part
  of speech.
• Partial matching not full matching
• Case 2
• w or one of its synonyms appears in the
  definition of the jth sense of w, D(w)j.
• Subcase 1
• w appears in D(w)j
• ? The sense of w is determined but the sense of
  w not.
• Subcase 2
• A synonym of the ith sense of w appears in D(w)j
• ? Both the senses of w and w are determined.

7
Case Analysis of Comparing WordNet Information

• Case 6
• A term t in the hyponym synset, H, of w appears
  in the definitions of several hyponym synsets of
  w.
• ? The sense of w is a hypernym synset of H.
• Case 8
• There are content words in common between the
  definition of a synset of w, and the definition
  of a hyponym synset of w.
• Case 9
• There is a common term in a hyponym synset of w
  and a hyponym synset of w.
• Case 11
• One or more senses of w and w belong to the
  same domain.

8
Case Conflict Resolution

• A term w may be assigned different sense due to
  the situation of satisfying several cases. ? case
  conflict
• The likelihood of choosing a sense of w is a
  function of 3 parameters
• Case weight the historical accuracy of the case
  in determining the sense of w.
• Sense weight the frequency of use of the sense
  of w.
• Supporting weight the historical accuracy of the
  case which determines the sense of w.

9
Case Weight

• CK-F and CK-P represent the full match and
  partial match cases, respectively, where K varies
  among the 11 cases.
• case_wt(CK-X)
• The weights of all cases are normalized so that
  the sum is 1.
• 5-fold cross validation.

10
Sense Weight and Supporting Weight

• The sense weight of a sense wi of term w is
• sense_wt(w, wi) f(w, wi) / F(w)
• where is the frequency of use of the sense wi of
  w and F(w) is the sum of the frequencies over all
  senses of w.
• Suppose w is disambiguated to sense wi using term
  w with sense wj
• sp_wt(wj)

11
The Disambiguation Function

• disam_wt(w, wi) sense_wt(w, wi)
• An example
• the queryhealth and computer terminal
• wterminal, wcomputer

12
Web Assisted Disambiguation

• The query is submitted to Google and the top 20
  documents are retrieved.
• For each document, find a window of n words, say
  50, which contains all query terms.
• All contents words in the window except w, form a
  vector.
• The vectors of the 20 documents are put together
  to be compared with the definitions of senses of
  w.
• The similarity function cosine

13
Experiment on WSD

• 250 queries in recent robust track of TREC
• There are 258 unique sense terms and 333
  ambiguous terms.

14
The Retrieval Model

• Noun phrases in queries are automatically
  identified and used for retrieval.
• The retrieval model considers that phrases are
  more important than individual words. Meng,
  2004
• Disambiguated query terms bring in new terms from
  WordNet.
• Pseudo-feedback is also used to expand query.
• Additional weights are assigned to feedback terms
  that are semantically related to the
  disambiguated query terms.

15
Experiment Result on IR

• Best-known results
• Overall 0.3333 Deng, 2004
• Hard 50 0.1941 Yu, 2004

16
Conclusion

• This paper provided an effective approach to
  disambiguate word senses in short queries.
• This approach can be applied to 100 of ambiguous
  terms and achieve an accuracy of 90,
  significantly better than any existing method.
• On average, the IR effectiveness of our system is
  7 better than the best result reported in the
  literature.