Active Taxonomies and their role in the Enterprise

1
Active Taxonomies and their role in the
Enterprise

• Jeremy Ellman
• School of Informatics,
• Northumbria University, UK
• Jeremy.Ellman _at_northumbria.ac.uk

2
Presentation Overview

• What is a Taxonomy?
• What are Enterprise Taxonomy Applications?
• How Taxonomies may be made Active using Lexical
  Chains
• The Case for Taxonomy
• Information Management and Metadata

3
What is a Taxonomy

• Minimally composed of a mostly hierarchical
  classification plus metadata
• Also called an Ontology
• Examples include
• Dewey Decimal System
• LoCSH
• (Local) Government Category List
• WordNet
• Rogets Thesaurus
• Entries may include synonyms, other language
  equivalents, author, date, related terms etc
• A working component of the information
  architecture

4
An Example Taxonomy
5
The Case for Taxonomy

• Taxonomies improve information access through
  structured browsing or search improvements
• Search process is limited by user behaviour
• Queries limited to 2.35 words on average
• Professionals achieve better results using more
  appropriate vocabulary
• Taxonomies allow synonyms to be added to user
  queries
• Heart attack Vs Myocardial infarcation
• Consistent use of metadata labels

6
Enterprise Taxonomy Applications and Function

• Applications include
• Content Management Systems
• Document Management Systems
• Portals and Intranets
• e-Commerce Applications
• Functions include
• supporting information retrieval
• Text classification
• Message Routing
• Topic Spotting
• providing the conceptual basis for knowledge
  based systems

7
Text Classification Methods

• Identify appropriate categories for a text
• Text is treated as Bag of Words
• Generally create a statistical model of each
  category derived from a training set
• Accuracy domain dependent
• At best 90
• Generally around 70

8
Key issues

• Current approaches to text analysis ignore
• text structure,
• Document meaning
• Semantics of categories

9
How Taxonomies may be made Active using Lexical
Chains

• An Active taxonomy would use the category
  information to assist processing
• Easy generation of navigation interfaces
• Rapid development of new content interfaces
• Reduced cost of localization
• Easy accommodation of name changes

10
What is a Lexical Chain?

• a set of words in a text that are related by both
  proximity, and by lexical linking relations
  between the words derived from an external
  knowledge source (Morris and Hirst 1991)
• Uses Rogets thesaurus for performance
• Simpler, more balanced structure than WordNet (4
  ply deep)
• 1000 main headings, divided into Noun, Verb ,Adj,
  Adv, and Phrase

11
Lexical Chains in text
Quotation from Einstein 1939 We suppose4 a very
long6 train0 travelling3 along the rails with a
constant velocity v and in the direction1
indicated in figure7 . People travelling in this
direction will with advantage5 use the train as a
rigid reference2 - body they regard all events
in reference-to to the train. Then every event
which takes place along the line also takes place
at a particular point10 of the train. Also, the
definition8 of simultaneity9 can be given
relative-to to the train in exactly the same way
as with respect to embankment
12
Actual Chains

• 0. train, rails, train, train, line, train,
  train, embankment,
• 1. direction, people, direction,
• 2. reference, regard, relative-to, respect,
• 3. travelling, velocity, travelling, rigid
• 4. suppose, reference-to, place, place,
• 5. advantage, events, event
• 6. long, constant
• figure, body

13
The Generic Document Profile

• Objective is to represent a document as taxonomy
  categories
• Document strength in these categories is derived
  from the lexical chains extracted
• Sum of Link Strengths weighted by link type and
  inter word distance
• Gives attribute value vector to represent
  document
• Can be compared to give measure of inter document
  similarity

14
GDP as Classification Method

• Advantages
• Gives list of categories ordered by value
• No training required
• Disadvantages
• Categories are those used in thesaurus or
  taxonomy
• Not all taxonomies have adequate vocabulary

15
Using the GDP As an Arbitrary Classifier

• Use machine learning training approach
• Trained documents produce profiles that can be
  merged, normalized, and associated with any
  category
• Classification process then compares against
  category models

16
Evaluation

• Compare against statistical approaches using
  reported corpora
• Initially using 20-newsgroups from BoW (McAllum
  (1999) 20Mb dataset
• Obtained WIPO Patent Dataset
• 450Mb XML zipped test data
• 500Mb XML zipped training data
• Also based on published Taxonomy

17
Research Questions

• Can the Bag of Words model be bettered with
  taxonomies?
• Can a subject specific taxonomy beat a generic
  one?
• Can subject specific taxonomies be extended
  accurately using generic ones?

18
Conclusion

• Taxonomies have a place in Enterprise Information
  Applications
• Text coherence can be used to characterise that
  text using external Active taxonomies
• This may be a practical text classification method

19
Questions?