Mining Anchor Text for Query Refinement

1
Mining Anchor Text for Query Refinement Reiner
Kraft and Jason Zien IBM Almaden Research Center
Mark Strohmaier
2
Problem Motivation

• 23 of search queries are single-term
• Expanding the query can lead to more accurate
  searches
• Previous studies indicate that anchor text is
  statistically similar to search queries
• Can this similarity be exploited to improve
  search queries?

3
What is anchor text?

• lta hrefthis is the websitegt This is the
  anchor text lt/agt
• Destination pages can have multiple links
  pointing to them
• Collections of anchor text can give a view of the
  destination page
• Naïve approach
• Find links whose anchor text is similar to the
  query
• Return the links destination pages to the user

4
Problems with naïve approach

• High term frequency is not directly related to
  page quality
• Repeated terms may lead to unnatural queries
• IDF is not necessarily relevant
• Anchor text may appear multiple times

5
Methods of Query Refinement

• Weighting the number of occurrences
• Weight based on the type of anchor text
• Number of terms in the anchor text
• Smaller terms is better
• Number of characters in the anchor text
• More concise queries are better

6
Benefits of the Anchor Text

• There is much less anchor text than document text
• Pages can have many incoming links
• Refined anchor text can capture a degree of site
  popularity

7
Mining Anchor Text

• Initial web crawl covered 33 million links on IBM
  intranet
• Additionally, roughly 350,000 queries were
  analyzed
• Both categories showed a similar relationship
  between length and number of occurrences

8
Pre-processing Summaries

• Query refinement is sensitive to the number of
  terms
• Too few may not lead to much improvement
• Too many may lead to overspecialization

Best results were for MAXCOUNT 3
9
Studies Performed

• Three different approaches were compared
• Anchor
• Ranked Anchor Text refinement
• Doc.SW
• This ranked pages based on the most frequently
  occurring 2 and 3 term phrases
• DOC
• Similar to Doc.SW, but not counting stop words

10
Ranking Anchor Texts

• The results are ranked based on
• WCOUNT score
• Number of terms in the anchor summary
• Number of characters in the anchor summary

11
Comparison of Methods

• Second comparison tested 22 different queries

• QUERYLOG processes and dynamically updates user
  queries based on previous ones, in a similar
  manner as ANCHOR

12
Conclusions

• Using anchor text leads to better results than
  performing similar methods on document
  collections
• A similar approach can be used to refine user
  search queries as well

13
Future Directions

• Broadening search queries
• Lexical analysis, rather than straight textual
• Pre- and Post- anchor text