Integration of Information Extraction with an Ontology

1
Integration of Information Extraction with an
Ontology

• KMi

• Knowledge Media Institute

• M. Vargas-Vera, J.Domingue, Y.Kalfoglou, E. Motta
  and S. Buckingham Sum

2
Introduction

• Ontology -gt Information Extractor
• English text (NLP)
• Group of tools their IE system
• KMi Ontology
• From UMass
• Marmot
• Crystal
• Badger
• OCML preprocessor

3
Presentation Layout

• Background on tool origins and area of work
• Description of tool integration
• Coping with ambiguity
• Description of output
• Population of Ontology
• Future Work

4
UMassUniversity of Massachutes Amherst

• Marmot, Crystal, Badger
• Classifies text by recognizing extraction
  patterns and semantic features associated to
  slots in predefined frames.

5
Testing Area KMi Planet

• Web-based new server
• Story Library
• Collections of news stories and postings
• Ontology Library
• Ontologies stored for use in extracting
  information from the story library.
• Uses OCML

• myPlanet

• myPlanet uses cue-phrases defined as research
  areas to query KMi planet through the ontology
  library and the information extraction tools
  were about to talk about

6
The Ontology Library

• 40 different types of events or activities that
  can be described by the ontology library.

• Event type 3 demonstration-of-technology
• technology-being-demostrated (technology) (Info
  Extraction)
• has-duration (duration) (30 min)
• start-time (time-point) (330pm)
• end-time (time-point) (4pm)
• has-location (a place) (room 120 TMCB BYU campus)
• other agents-involved (list of person(s)) (Dr.
  Embley)
• main-agent (list of person(s)) (Brian Goodrich)
• location-at-start (a place) (room 120 TMCB BYU
  campus)
• location-at-end (a place) (room 120 TMCB BYU
  campus)
• medium-used (equipment) (mutli-media projector,
  ppt)
• subject-of-the-demo (title) (Integration of
  Information Extraction with an Ontology)

7
Marmot

• Natural Language Processor
• Noun, Verb, and Prepositional Phrases
• John Domingue Wed, 15 Oct 1997.
• David Brown, University for Industry visits the
  OU.

• ltexgt 2 1
• SUBJ(1) DAVID BROWN COMMA UNIVERSITY
• PP (2) FOR INDUSTRY
• VB (3) VISITS
• OBJ1(4) THE OU
• PUNC(5) PERIOD
• lt/exgt

• ltexgt 1 1
• SUBJ(1) JOHN DOMINGUE
• ADVP(2) _at_WED_COMMA_15_OCT_1997_at_
• PUNC(3) PERIOD
• lt/exgt

8
Crystal

• Dictionary Induction Tool
• Using keyword to annotate text with semantic
  tags.
• Visitor (ltVIgt David Brown ltVIgt)
• Place (ltPLgt the OU ltPLgt)
• Specific-to-general driven data search
• Relaxes constraints on initial definitions until
  it finds the most specific
  definition that covers all
  instances of the word in the text.
• Retains results for future use
• Tested on over 300 stories, 100 precision
  and recall

9
Badger

• (fairly certain whoever wrote this section did
  not speak English as first language)

Matches sentences from text against concept nodes
passed from Crystal. Select the best match by
max number of features matching the concept
node. Can remove irrelevant sentences from
problem set.

• gt

• http//rockape.qgl.org/crap/badger.swf

10
Coping with Ambiguity

• Query list of institutions

• Return list of institutions no match

• Query list of projects

• Return list of project - match

• No discussion of whether this was automatically
  done by the extractor or manually by the users.

11
OCML Code Translator (Operational Conceptual
Modeling Language)

• Tokenise Badger output, find corresponding CN
  definitions and extract all the objects found in
  the story

12
Ontology Maintenance

• Use Badger (lexicon) and Crystal (concept) output
  to automatically update Ontology library whenever
  a new story is added to the Story library
• Some cannot be automatically updated
• There is not enough information in the story
• No current template to match with the sentence
  concepts.

13
Conclusion

• IE system created using Marmot, Crystal, Badger
  and the OCML translator.
• Obtained good results in KMi stories.

Assessment

• Sporadic periods of quality technical writing,
  interspersed with nearly impenetrable English
• A borrowing of tools, translated to OCML and
  ported for KMi

14
Future Work

• Deriving the type of an object when it does not
  match a predefined template.
• Automatic creation of new classes and subclasses.
• Using this IE tool in other domains (need new
  training data?)
• Trying out a new Machine Learning algorithm in
  Crystal and comparing performance.
• Using the IE tool hypertext.
• Saving Badgers output in XML
• Creating a more visual gui for the ontologies.