Collaborative Query Profiling

1
Collaborative Query Profiling

• Erik W. Selberg
• University of Washington
• Computer Science and Engineering

2
Outline

• HuskySearch
• The Vocabulary Problem
• The Indexing Problem
• Collaborative Query Profiling
• Conclusions and Future Work

3
HuskySearch

• Parallel Search Engine
• Send query to all main services
• AltaVista, Lycos, WebCrawler, etc.
• Collate results into single list
• Post-process results for quality
• Key Idea Dont use one, use all!

4
The Vocabulary Problem

• Semantically equivalent search phrases dont
  yield the same results
• Different search terms yield different results
• Different people use different search terms for
  the same question
• People expect precise and accurate answers

5
Example

• What were Princes last five albums, and in what
  year were they released?
• prince
• discography prince
• prince albums
• the artist formerly known as prince
• prince album release dates

6
The Indexing Problem

• Which terms should be used for indexing?
• Hint All of them is a bad choice
• Which documents are indexed?
• Is there higher level information?

7
Example

• Garbage terms - typos, code, etc.
• Stop words
• Lycos index only 100 relevant terms per page
• AltaVista finite global list of indexable terms
• WebCrawler search to depth 3

8
HuskySearch
9
prince discography
10
albums Prince
11
prison us citizen charge
12
US citizens foreign jails
13
Collaborative Query Profiling

• Helping users find relevant documents with input
  from past queries
• How are users helped?
• Which queries are useful?
• What are the metrics?

14
Implementation

• Rank function twiddling augment results
• Relevance feedback
• Create new searchable indices
• Include indices with HuskySearch queries
• Adds new pages to results list
• Augments pages returned from other sources
• Evaluate using passive testing

15
Indexed Results

• Observation Snippets in HuskySearch results
  pages highlight relevant terms
• Experiment 1 Create an index of result pages
  and search them
• Experiment 2 Create an index of good result
  pages and search them
• Note this creates an implicit searchable query
  history

16
Collaborative Databases

• Hypothesis pages close to user searches create
  a good web snapshot
• Exp. 3 Create an index of pages referenced in
  results
• Exp. 4 Create an index of references pages
  clicked on
• Exp. 5, 6 Expand 3, 4 to all pages within 3
  links

17
Metrics

• Clicking implies relevance
• Clicking top ranked results implies validation
• Or are we just measuring what weve trained users
  to do?
• Need user testing to confirm these metrics

18
Current Status

• Exp. 1 already integrated
• Initial pilot study
• Exp. 3, 5 coming quickly
• Exp. 2, 4, 6 soon
• Followup user study planned

19
Conclusions

• Vocabulary and Indexing Problem make it harder to
  find information
• Collaborative Query Profiling aids users with
  previous searches
• Integrated databases fit into user interface
  nicely
• Collaborative aspect allows users to leverage
  other users intelligence