An Artificial Immune Based Approach to Semantic Document Classification

1
An Artificial Immune Based Approach to Semantic
Document Classification

• Julie Greensmith Steve Cayzer
• BICAS research group
• Hewlett-Packard Laboratories, Bristol
• (University of Leeds)
• julie.greensmith_at_hp.com
• September 2003

2
The Problem Information Overload

• Wealth of Internet resources
• Googlefact August 2003, searches approximately
  3 BILLION PAGES
• Drowning in data
• Need to add structure to the information

3
Current Solutions

• Use metadata to describe meaningful relationships
• Yahoo pages organised into a taxonomy

MUSIC
POPULAR
CLASSICAL
DANCE
ROCK
CLASSICAL
BAROQUE
4
HOWEVER

• Requires manual annotation i.e. effort
• Markup of both new and existing pages
• Placing of documents into taxonomies subjective
• Need a personalised system to add structure

5
So, what should we do about it?

• Automatic Documents Classification
• Less labour intensive
• Multiple Taxonomy Mapping
• Number of potential approaches
• Decision tree
• Naïve Bayesian
• Use a more novel computational approach namely
  an artificial immune system as a document
  classifier

6
AIRS and Document Classification

• AIRS
• Developed by Andrew Watkins
• B-Cell inspired, resource limited algorithm
• K-nearest neighbour
• Supervised learning paradigm
• Multi-class classification
• Good performance on other data mining tasks
• What we would like to achieve
• Push AIRS into the novel domain of document
  classification, and more significantly, semantic
  document classification

7
So, how are we going to achieve this?

• Creation of feature vectors
• Extraction simple text processing
• Selection information gain
• Representation boolean feature representation
• Present feature vectors to classifier system
• Measure predictive accuracy
• Hierarchical placing of documents
• Multiple Taxonomy mapping
• Placement in taxonomy to derive semantic features
• Semantic similarity measures
• Personalised Semantic Structure

8
Work performed so far

• Document classification
• Yahoo taxonomies business pages
• 2 classes
• Small, medium large datasets
• Multi-class datasets
• Varied amount of features
• Comparison with a naïve Bayesian based system
• Preliminary Results
• Promising ?
• Presented in my MSc Thesis

9
Some Preliminary Figures

• No significant difference
• found between 2 class and
• 4 class dataset

• AIRS performed significantly better than
  alternative naïve Bayesian system

10
Semantic Antics

• The AIRS system can classify documents
• New frontiers for an artificial immune system
• Perform multiple taxonomy mapping for
  personalisation
• Use semantic features in conjunction with text
  features to develop hybrid feature vectors
• Semantic similarity measures
• This is work to be performed in the near future ?

11
Summary

• Suffering from data overload
• Alleviate the solution with hierarchical document
  classification
• Use AIRS as a document classification tool
• Achieved promising results on a number of
  datasets
• Use AIRS as a semantic document classification
  tool
• Semantic similarity metrics
• Personalised semantic structure

12
(No Transcript)