Relationship-Based Top-K Concept Retrieval for Ontology Search

Relationship-based Top-K Concept Retrieval for Ontology

Search

Anila Sahar ButtAnila Sahar Butt, Armin Haller, Lexing Xie, Armin Haller, Lexing Xie

The Australian National UniversityThe Australian National University

[email protected]@anu.edu.au

2

Motivation – Ontology Search

“An ontology is a formal, explicit specification of a shared conceptualization.” [Gruber 1992]

A central ingredient for effective ontology re-use is the discovery of the

“right” ontology or ontological term for a use case

Motivation – Ontology Search

• Ontology Search– Matching a search term with a more

expressive class description

• Matching terms are defined with differing– Perspectives– Levels of detail– Reuse and Extensions

3

How to rank the similar concepts with different levels of modelling

detail?

4

Relationship-based Top-k Concept Retrieval

• The framework retrieves top-k concepts for keyword query

DWRank - Ranking Model Top-k Filter

5

DWRank – Dual Walk Ranking Model

6

DWRank – Dual Walk Ranking Model

For a simple keyword query: Rank of a concept

Semantic Similarity: Text relevancy of the concept Coverage : Centrality of the concept Reuse : Authoritativeness of the Ontology

DWRank – Dual Walk Ranking Model

1. Query independent scores for each concept of ontologies based on their importance

Centrality of the concept - HubScore Authoritativeness of the Ontology - AuthScore

1. Relevance score of a concept to a query:

• DWRank Function: Linear model combines – Text relevancy of the concept description to a query– HubScore and AuthScore

8

HubScore – Centrality of a Concept

Connectivity : Relations starting from the concept

Neighbourhood :

Relations starting from the concept to another central concept

9

@prefix a: http://example.org/def/people#

a:Person

a:Organization a:Project

0.14

0.46

0.380.26

0.140.14

0.14

0.14

Reverse PageRank

AuthScore – Authoritativeness of an OntologyReuse : Relations ending at the ontology

Neighbourhood : Relations starting from another authoritative ontology to the ontology

10

:People:Restaurant:Location

0.4710.1450.10

PageRank

Zubeida

Zubeida Zubeida

Zubeida

11

DWRank Function

• The Ranking model is function of:• Concept Text Relevancy• HubScore• AuthScore

∑∈

=+=

Qq

vssv

*2 *1vO)(v,

))φ(q(q,fQ)(v,F

a(O)]w O)h(v,[w *Q)(v,FR

12

DWRank Score

Query: Persono Fv(v,Q) = 1o h(v,O) = 0.46o a(O) = 0.471

a(O)]w O)h(v,[w *Q)(v,FR *2 *1vO)(v, +=

@prefix a: http://example.org/def/people#

a:Person

a:Project

0.14

0.46

0.380.26

0.140.14

0.14

0.14

0.471

0.466

0.471]0.5 0.46[0.5 *1 * *

=+=

13

DWRank vs. Linked-based Ranking Models1. Direction of the walk varies based on the

link type Intra-ontology links: Reverse PageRank Inter-ontology links: PageRank

14

DWRank vs. Linked-based Ranking Models (cont’d)

2. Linked Analysis : HubScore – Concept

o Independently on each ontology AuthScore – Ontology

o Ontology Corpus

Top-K Filter

15

16

Intended Type Filter

• Intended Type vs. Context Resource Name of the Person

o Intended Type: Nameo Context Resource: Person

Relationship-based Top-k Concept Retrieval Phases

17

Relationship-based Top-k Concept Retrieval

• The framework retrieves top-k concepts for keyword query

– Offline Ranking and Index Construction– Online Query Processing

18

Offline Ranking Index Construction

19

Offline Ranking Index Construction

20

Offline Ranking Index Construction

21

Offline Ranking Index Construction

22

Online Query Processing

23

Candidate Result-set Selection

Candidate Result-set Selection

HubScore and AuthScore Selection

HubScore and AuthScore Selection

Relevance Score of CandidateList and

Ordering

Relevance Score of CandidateList and

Ordering

Intended Type FilterIntended Type Filter

Ontology Corpus

Ontology Corpus

IdxIdx

HubConIdxHubConIdx

AuthOntIdxAuthOntIdx

UserQuery

Results

Evaluation

• Effectiveness of the approach– Two versions of framework

• DWRank • DWRank + Filter

• CBRBench – CanBeRra Ontology Benchmark– Ten sample queries– Human evaluated gold standard– Baseline Ranking models

24

Evaluation (cont’d)

• Effectiveness metrics – Precision @ k – Mean Average Precision @ k – Discounted Cumulative Gain @ k – Normalized Discounted Cumulative Gain @ k

25

DWRank Effectiveness

Intended Type Semantics Filter Effectiveness

Conclusion & Future Work

• We presented– Ontology Search– Framework for top-k concept retrieval– DWRank- Dual Walk Ranking Model– Experimental Evaluation

• Ranking ontologies for compound concepts

28