Search with Context

1
Search with Context

• Reiner Kraft and his colleagues, Yahoo! Research
• CIKM05 Y!Q Contextual Search at the Point of
  Inspiration
• WWW06 Searching with Context

2
Outline

• Motivation
• Algorithm
• Evaluation

3
Motivation

• User queries often convey ambiguous and
  incomplete user information need
• Context can help clarify user information need
• Capturing context (Y!Q, CIKM05)
• Exploiting context (WWW06)

4
Y!Q

• Demo of Y!Q
• Contextual search engine (server side)
• JavaScript at front-end for user interaction
  (client side), like IntelliZap

5
Y!Q Architecture
6
Three Modules of Y!Q

• Content Analysis (CA)
• Select terms from context and queries
• Query Planning and Rewriting Framework (QPR)
• Determine Service Providers (SE, News)
• Rewrite queries
• Contextual Ranking (CR)
• Aggregate search results

7
Overview of WWW06 Paper

• Algorithms to do contextual search
• Query Rewriting (QR)
• Rank Biasing (RB)
• Iterative Filtering Meta-search (IFM)

8
Algorithm Description
9
Query Rewriting and Rank Biasing

• Query Rewriting
• Use standard search engine (Simple)
• Parameter terms in contextual query
• Rank Biasing
• Special search engine (rank biasing operator)
• ltquerygt ltselectioncatgtltoptionPersian, 2.0gt

10
Iterative Filtering Meta-search

• Query generation sliding window
• Ranking aggregation
• Ranking average
• Markov Chain
• States are URL
• If average URL u rank higher than URL v, there is
  a link from u to v
• Small transition probability from every url to
  every other url (ergodic and aperiodic)

11
Evaluation

• Dataset 200 context benchmark collected from the
  logs
• Relevance judgment 24,566 judgments from 28
  judge expert
• Evaluation metrics P_at_1, P_at_3, SP_at_1, SP_at_3

12
Experiment Procedure

• A judge selects context (most or least
  comfortable) from the benchmark
• For each web search result, four choices 1)
  relevant, 2) somewhat relevant, 3) no, 4) cant
  tell

13
Major Findings

• QR performs quite well
• IFM performs well in recall and relevance
• Human manually queries are not good enough

14
Experiment Results Query Rewriting
15
Experiment Results (cont.)
16

• The End