Applying a Lexical Similarity Measure to Compare Portuguese Term Collections

1
Applying a Lexical Similarity Measure to Compare
Portuguese Term Collections
Programa de Pós-Graduação em Ciência da
Computação Faculdade de Informática Pontifícia
Universidade Católica - RS

• Marcirio Silveira Chaves
• Vera Lúcia Strube de Lima

2
Context I

• Ontological Structures and Ontologies
• Reuse of Knowledge
• Semantic Web

3
Context II

• Mapping of OSs
• Similarity between terms (distinct matizes)
• Similarity Measures
• Edition Distance
• Two levels of similarity
• The Portuguese language

4
Summary

• Context
• Research Question
• Similarity Measures
• Lexical Similarity Measure
• Experiments
• Final Remarks and Future Work

5
Research Question

• How to map similar concepts between distinct
  ontological structures?
• Hypothesis
• There is a degree of similarity among ontological
  structures independently created, which can be
  detected in order to allow a mapping.

6
Similarity Measures I

• Lexical Level
• Edit Distance (Levenshtein 1966) - ED
• The minimum number of
• insertions,
• deletions or
• substitutions (reversals)
• necessary to transform one string into another
  using a dynamic programming algorithm.
• e.g. ED (ocidente, oriente) 2

7
Similarity Measures II

• Lexical Level
• String Matching (Maedche e Staab 2002) - SM
• e.g.

8
Lexical Similarity Measure I
k reaches the amount of words of the term with
the minimum number of words.
9
Lexical Similarity Measure II

• Example

10
Characteristics of the Experiment

• Multidomain Experiment
• OSs from
• Brazilian Senate Thesaurus (OSA)
• São Paulo University - USP Thesaurus (OSB)
• single-word and multiword terms
• validation and evaluation
• Terms in OSA categorized into two sets for each
  phase
• Terms in OSB remained without categorization
  during both phases.

11
First Letter Heuristic
12
Relevant Numbers

• Evaluation phase
• single-word terms
• 1,823 from OSA
• 7,039 from OSB
• multiword terms
• 4,701 from OSA
• 16,986 from OSB.
• 2,887 pairs of terms similar (using SM or LS)

13
Evaluation

• Analysis of Results

14
Analysis of Group G5

• Most of the pairs analyzed during the evaluation
  phase (73)
• 907 single-word terms
• 1,211 multiword terms.
• Extract of group G5 single-word and multiword
  terms

15
Same Domain Experiment I

• GEODESC Thesaurus (2,083)
• USP Thesaurus (429 terms ) belonging to the
  Geosciences domain.

16
Same Domain Experiment II

• Analysis of Group A

Pairs of terms considered dissimilar by SM and
similar by LS
17
Same Domain Experiment III

• Analysis of Group A

Pairs of terms considered similar by SM and LS
18
Same Domain Experiment IV

• Analysis of Group B

Multi and single word pairs of terms
19
Same Domain Experiment V

• Analysis of Group B

Contribution of the penalties without first
letter heuristic
20
Final Remarks and Future Work

• About this work
• Creation, validation and evaluation of LS measure
• One of the first efforts to deal with Portuguese
  term collections
• Experiments with terms belonging to multidomain
  as well as to the same domain structures
• Future Work
• Application of LS measure to other languages
• Treatment of semantic-structural level in the
  structures