Modelling and Use of DomainSpecific Knowledge for Similarity and Retrieval

1
Modelling and Use of Domain-Specific Knowledge
for Similarity and Retrieval

• Troels Andreasen, Henrik Bulskov, and Rasmus
  Knappe

2
Outline

• Motivation
• Representation of Ontologies
• Modelling Ontologies
• General Ontology
• Domain-specific Ontology
• Deriving Similarity
• Applications
• Prototype System
• Navigating and Surveying
• Query Visualization

3
Motivation

• The use of ontologies
• Contributes to organization of concepts,
  structure and relations within a knowledge domain
• Which in turn
• Provides means for enhanced, knowledge-based
  approaches to surveying, indexing, and querying
  of document collections

4
Concept Language - Ontolog

• Basically a lattice-algebra with attribution
• Compound concepts are built from
• atomic concepts of the ontology, and
• attribution
• attribute features using semantic relations like
• WRT with respect to
• CHR characterized by
• CBY caused by
• PNT patient of act or process
• LOC location, position
• ...

5
Representation of Ontologies

• A generative framework consisting of
• A basis ontology Obasis A,ISAKB,
• A generative concept language (OntoLog) defining
  the set of well-formed concepts L. Given atomic
  concepts A and a set of semantic relations R, the
  set of well-formed concepts L are

6
Concept Examples

• The red and yellow bird
• birdCHRred, CHRyellow
• The meeting on Thursday
• meetingTMPthursday
• Disorder caused by lack of vitaminC
• disorderCBYlackWRTvitaminC

7
Modelling Ontologies

• Modelling in this context consists of two parts
• The inclusion of knowledge from available
  resources into a general ontology, and
• A restriction to the part of the general ontology
  covering the instantiated concepts in the
  document collection

8
The General Ontology (1)

• Built from various sources
• Taxonomy supplemented with word and term lists,
  dictionaries, thesauri
• We assume the presence of taxonomy in the form of
  a simple taxonomic concept inclusion relation
  ISAKB over the set of atomic concepts A

9
The General Ontology (2)

• Based on ISATRAN the transitive closure of ISAKB
  we generalize into a relation over well-formed
  concepts L
• where repeated in each inequality denotes zero
  or more attributes of the form

10
The General Ontology (3)

• Ogeneral L,,R thus encompasses
• A set of well-formed concepts derived in the
  concept language from a set of atomic concepts A
• An inclusion relation generalized from a
  knowledge expert given relation ISAKB
• A supplementary set of semantic relations R

11
Domain-specific Ontology

• Transform the global generative ontology into a
  domain-specific ontology
• Restrict the global ontology to cover only the
  concepts in a collection for instance those
  appearing in a given document collection.
• The result can be considered as an instantiated
  ontology with respect to the document collection

12
Domain-specific Ontology,an example
Example knowledge base ontology ISAKB
Example instantiated ontology
13
Deriving Similarity

• Domain-specific ontology as basis for deriving a
  similarity measure for use in connection with
  querying of documents
• The measure must reflect nearness of concepts in
  the ontology
• Well-known approach Shortest Path
• Problem Multiple connections are ignored
• Applied here
• Shared Nodes

14
Shortest path Similarity

• Over ISA-relations

sim(dogCHRgray , catCHRgray )
sim(dogCHRgray , birdCHRyellow)

• Counterintuitive
• dog and cat share the property gray

15
Shared Nodes

• Similarity between two concepts in this approach
  is based on the set of upwards reachable nodes
  shared between the concepts

16
Shared Nodes Similarity

• Over all relations

sim(dogCHRgray , catCHRgray ) gt
sim(dogCHRgray ,birdCHRyellow))

• Notice
• Shared nodes take multiple connections into
  account

17
Shared Nodes Similarity

• with as the nodes upwards reachable from x
• similarity is proportional to
• We have
• where the triple (x,y,r) is the edge of type r
  from x to y, E the set of all edges, T the
  topmost concept, and r ? R ISA, CBY, CHR,

18
Shared Nodes Similarity
sim(dogCHRgray , catCHRgray ) gt
sim(dogCHRgray , dogCHRlarge )

• Counterintuitive
• concept-inclusion (ISA) should have higher
  importance than characterized-by (CHR) property

19
Weighted Shared Nodes Similarity
sim(dogCHRgray , catCHRgray ) lt
sim(dogCHRgray , dogCHRlarge )

• solution
• attach weights in 0,1 to relations so the
  nodes upwards reachable from x
  becomes a fuzzy set

20
Weighted Shared Nodes Similarity

• Similarity is proportional to
• Shared nodes
• Weighted Shared Nodes
• where the function weight(r) attach a weight to
  each relation type.
• We assume that
• weight(ISA) 1
• and the weight attached to all other relations
  are less than 1.
• For instance
• weight(CHR) 0.8

21
Similarity measure

• Similarity can be defined in various ways,
• An asymmetric measure
• is used here where contributes to a
  weighted average that determines the degree of
  influence of the nodes reachable from x
  respectively y.

22
Applications

• Prototype system
• Navigating and Surveying
• Query Visualization

23
Prototype System

• Based on
• WordNet 2.0
• Suggested Upper Merged Ontology (SUMO)
• Mid-level Ontology (MILO)
• Contains
• Approximately 100.000 concepts (synsets)
• ISAKB hypernym relation

24
Navigating andSurveying

• Instantiated Concepts
• palisade,
• stockadechrold,
• rampartchrold, and
• churchchrold
• Two aspects
• Fortifications
• Place of worship
• More abstract
• Buildings
• Something dated back in time

25
Visualizing Queries

• Polysemous concepts
• Q bank,huge
• Domain specific concepts

26
Conclusion

• Domain-specific ontologies
• Restriction of the general ontology, has
  application with respect to
• Deriving Similarity
• Weighted Shared Nodes
• Navigation and Surveying
• Query Visualization