Clustering Top-Ranking Sentences for Information Access

1
Clustering Top-Ranking Sentences for Information
Access

• Anastasios Tombros, Joemon Jose, Ian Ruthven
• University of Glasgow University of Strathclyde
• Glasgow, Scotland

2
Some Background Motivation

• Challenge How to provide effective access to
  information
• Approach Combine clustering top-ranking
  sentences (TRS)
• clustering has been used extensively on the
  document level
• TRS are based on single document summaries
• Overall aim of the work
• to create a personalised information space
• to use information from users interaction

3
Top-Ranking Sentences

• Assume a user with a query
• the query is sent to an IR system
• consider only the top retrieved documents, e.g.
  30
• apply a query-biased sentence extraction model to
  each of these documents
• construct a sentence extract of max. 4 sentences
  per document
• the set of these sentences for the 30 documents
  is the set of TRS
• TRS can be ranked by their query-biased scores

4
Top-Ranking Sentences (cntd.)

• TRS have shown to be effective in interactive IR
  on the Web
• they provide effective access to the retrieved
  information
• They can be seen as a level of abstraction of the
  set of retrieved documents
• We introduce an extra layer of abstraction by
  clustering the set of TRS

5
Clustering Top-Ranking Sentences

• An attempt to create a personalised information
  space
• sentences give local contexts in which query
  terms occur
• sentences discussing query terms in similar
  contexts should cluster together
• this structure should facilitate a more intuitive
  and effective access to information
• Similarities and differences to document
  clustering

6
Comparing TRS and Document Clustering

• We used 4 searchers with a total of 16 queries
• each searcher assessed the utility of the top 30
  documents on a scale of 1-10
• For each query
• we downloaded the top-30 retrieved documents
• we extracted the set of TRS
• we clustered the 30 documents and the set of TRS
• we assigned scores to document TRS clusters
• sum of the document (sentence) scores divided by
  the number of documents (sentences) in the cluster

7
Some Results

• Scores of TRS clusters were significantly higher
  than those of document clusters
• best cluster averages 4.78 vs. 5.82
• overall averages 3.2 vs. 3.73
• Average precision and recall were higher for TRS
  clusters
• define P R based on documents with scores 7
• average P 0.38 vs. 0.49
• average R 0.73 vs. 0.77
• Cluster sizes were comparable
• 5 docs per cluster vs. 5.3 sentences per cluster

8
Conclusions Future Plans

• TRS clusters have the potential to offer more
  effective information access
• only one aspect of their expected utility
• Integrate TRS clustering in interactive web
  searching
• investigate its utility in user-based studies on
  the live Internet
• We have extended the reported work
• more searchers queries, different clustering
  methods
• inter-sentence similarities, structure of
  information space