WordSieve: Learning Task Differentiating Keywords Automatically

1
WordSieve Learning Task Differentiating Keywords
Automatically

• Travis Bauer
• Sandia National Laboratories
• (Research discussed today was done at Indiana
  University)

2
Learning Task ContextsCalvin

• Learn what characterizes a users task contexts
• Unobtrusive Observing
• Keyword Extraction
• Index based on Context

3
Currently Used Algorithms

• TFIDF
• Latent Semantic Analysis
• Log-Entropy

4
Currently Used Algorithms

• TFIDF
• "One of the most successful and well tested
  techniques in Information Retrieval." - Pazanni
• Syskill Webert (Pazanni '96)
• Hierarchical Feature Map (Merkl '97)
• Learning in Document Filtering (Callen '98)
• Topic Detection (Shultz '99)
• Remembrance Agent (Rhodes '00)
• Lexical Signatures (Park '02)
• Latent Semantic Analysis
• Log-Entropy

5
Currently Used Algorithms

• TFIDF
• Latent Semantic Analysis
• Well known, popular, well covered in the
  literature
• Grading Essay Tests
• Taking Physics tests
• Taking synonym exams
• Cross Linguistic IR (Dumais '97)
• Assigning papers for peer review (Dumais '92)
• Information Filtering (Foltz '90)
• Log-Entropy

6
Currently Used Algorithms

• TFIDF
• Latent Semantic Analysis
• Log-Entropy
• Not used as much for Personal Information
  Retrieval
• Higher overhead than TFIDF
• Indexes based on the distribution of terms across
  documents potentially better performance

7
Comparison to Current Techniques

• Current Techniques
• Static Corpora
• Comprehensive Statistics

• WordSieve
• Neural Network-like processing
• Stream of data
• Local learning
• Competitive Learning

8
Good Discriminator of Context
9
WordSieve Concept
User Browsing
Attributes Term Activation Priming
10
WordSieve 1
Words Absent in Document Sequences
User Profile
Context Profile
Words Occurring in Document Sequences
Words Currently Occurring Frequently
11
WordSieve 2
User Profile
Words Reflecting Context
Context Profile
Words Currently Occurring Frequently
12
Web Browsing Data Set

• Sixteen Users
• Four Topics, 10 minutes Each
• Political Life Al Gore
• Political Life George Bush
• Traditional Indonesian Cooking
• Traditional Thai Cooking

Categorized Document Set
Automatically Generated Queries
13
Browsing Results
14
Contributions

• It is possible to extract context differentiating
  terms from document streams using unsupervised
  competitive learning.
• Comprehensive statistics are not necessary in the
  described situations given an ordering of the
  documents.
• Performance is comprable to LSI and better than
  Log-Entropy and TFIDF

15
Potential Next Steps

• WordSieve
• Automate Parameter Optimization
• Co-occurrance of terms
• Other Domains
• Multi-dimensional data stream
• Machine Vision

16
Support

• This work was conducted under the advisement of
  David Leake at Indiana University.
• It was sponsored in part by the GAANN fellowship.
• The original version of the personal information
  agent was designed and written with partial
  support from NASA under award No NCC 2-1035

17
For More Information

• Travis Bauer
• www.cs.indiana.edu/trbauer/publications.htm

18
Usenet Data Set

• Three sets of 5 newsgroups
• alt.atheismtalk.religion.miscsoc.religion.christ
  ianrec.sport.baseballrec.sport.hockey
• comp.os.ms-windows.misccomp.sys.ibm.pc.hardwarec
  omp.sys.mac.hardwarerec.autosrec.motorcycles
• talk.politics.gunstalk.politics.miscsci.electron
  icssci.medsci.space

Categorized Document Set
Automatically Generated Queries
19
Usenet Results