Similarity Measures for Query Expansion in TopX

1
Similarity Measures for Query Expansion in TopX

• Caroline Gherbaoui

Universität des Saarlandes Naturwissenschaftlich-
Technische Fak. I Fachrichtung 6.2 - Informatik
Max-Planck-Institut für Informatik AG 5 -
Datenbanken und Informationssysteme Prof. Dr.
Gerhard Weikum
2
Overview

• background knowledge
• similarity measures for the query expansion
• evaluation of the computed similarity values
• changes in TopX
• conclusion

3
Background

• top-k query processing
• provides k most relevant results
• query expansion
• extends source query terms
• word sense disambiguation
• extracts correct meaning
• ontology
• amount of terms with their meanings and semantic
  relations

4
Word Sense Disambiguation
java, coffee
island
coffee
java
programming language

5
Query Expansion
COFFEE
drink, espresso
6
TopX

• top-k retrieval engine
• text and XML data
• word sense disambiguation
• query expansion
• ontology

7
TopX WordNet Ontology

• lexicon for the English language
• hierarchical relations
• one relation ? one direction
• 160,000 words
• 120,000 synsets
• 210,000 relations

8
TopX YAGO Ontology

• Wikipedia and WordNet
• hierarchical and not hierarchical relations
• one relation ? two directions
• 2,100,000 words
• 2,200,000 concepts
• 6,000,000 relations

9
Similarity Measures

• Dice similarity
• the already used measure in TopX
• NAGA similarity
• applied measure for YAGO
• Best WordNet similarity
• measure with best result among WordNet measures

10
Dice Similarity Measure

• sdfsdf
• measures the intersection of two regions

11
NAGA Similarity Measure

• sdfasfsdf
• combination of the confidence of a relation and
  the informativeness of a relation

12
Best WordNet Similarity Measure

• sdfsdfsdf
• product of the transfer function of the path
  length and the transfer function of the concept
  depth

13
Evaluation
14
Evaluation

• DICE measure ? applicable
• also on the YAGO ontology
• NAGA measure ? applicable
• with omitting of the forward direction
• Best WordNet measure ? not applicable
• due to the density of YAGO

15
Changes for TopX

• tuning of some procedures
• Dijkstra algorithm
• word sense disambiguation
• query expansion
• extension of configuration file

16
Conclusion

• larger knowledge base
• more flexibility
• increased complexity
• further measure for the similarity computation ?
  NAGA similarity

17

• Questions?