QuerySensitive Similarity Measures for Information Retrieval

1
Query-Sensitive Similarity Measures for
Information Retrieval

• By Anastasios Tombros and C.J. van Rijsbergen
• Presented by Brad Wardman

2
Introduction

• Cluster analysis a technique that allows the
  identification of groups, or clusters, of similar
  objects in a space that is typically assumed to
  be multidimensional
• The central idea is that closely associated
  documents tend to be relevant to the same requests

3
Alternative View of Cluster Hypothesis

• The hypothesis should not be seen as a test for
  an individual collections clustering tendency,
  but the hypothesis should be valid for every
  collection

4
Hypothesis

• For any given query, pairs of relevant documents
  will exhibit an inherent similarity that is
  dictated by the query itself

5
Query-Sensitive Similarity Measures

• Biases interdocument relationships toward pairs
  of documents that jointly possess attributes that
  are expressed in a query
• The query terms are the salient features that
  define the context under which similarity of two
  documents is judged

6
Static clustering

• Clustering has already been applied statically
  over entire document collections prior to
  querying
• The value of the similarity will stay the same
  under all queries that a user may pose on an IR
  system

7
Cosine Coefficient

• Commonly used as a measure of interdocument
  relationships

8
Hearst and Pederson

• Postulated that relevant documents tend to appear
  in the same clusters, but the clusters are
  created as a f(x) of the documents retrieved in
  response to a query therefore, have the
  potential to be more closely tailored to the
  characteristics of the query than a static
  clustering

9
Information Retrieval Goal

• The goal is, for any query, to place relevant
  documents closer to each other than nonrelevant
  ones
• Interdocument similarity is dynamic and changes
  explicitly depending on the query

10
QSSM Model - M2

• Q query vector
• Di,Dj document vector
• C Di n Dj vector containing common terms of
  the document vectors

11
M3

• Linear combination of previous two models
• Pairs of documents with more terms in common with
  the query than other pairs will be assigned
  higher similarity values

12
M1

• Product of the two sources of information
• Assumes that the presence of query terms is
  required for a document to be relevant

13
Limitations

• Works on the assumption that query terms are
  sufficient indicators of document relevance
• Only takes one instance of users information
  need into account
• Short queries (2-3 term queries)
• It will be doubtful whether the similarity
  measures will have enough information

14
Results

• M1 and M3 are significantly more effective than
  the cosine coefficient at placing co-relevant
  documents closer to each other
• Which provides a greater chance for a more
  effective clustering of the document space
• M1 and M3 achieved higher scores than M2 for
  almost every condition

15
Results Cont.

• The lower effectiveness of M2 is not surprising
  since given that this measure uses less
  information than the other two measures
• M2 is highly affected by query length
• M1 and M3 do not seem to be affected by query
  length
• The difference between M1 and M3 are not
  statistically significant

16
Conclusions

• Similarity is a dynamic and purpose-sensitive
  notion
• QSSM have the potential to capture the dynamics
  of similarity for the calculation of
  interdocument relationships