Semantic similarity measures for information retrieval systems using ontology

Abstract

This work focuses on devising computational models for assessing similarity among words/concepts in the knowledge sources like ontologies. Semantic similarity assessment plays an important role in the fields of Psychology, Information Retrieval and Information Integration systems. The paradigm shift of syntactic web to semantic web has emphasized the use of development of semantic similarity measures to computationally identify related concepts within and among ontologies. This work deals with semantic similarity approaches which exploit the concept relationships associated with the concepts to quantify similarity among concepts defined within and among ontologies. The work specifically is interested in proposing corpus independent information content based measures for quantifying similarity among concepts belonging to single and multiple knowledge sources. The information content computation of these measures has been redefined with taxonomic and non taxonomic relations possessed by the concepts defined in the ontology. This new definition of information content solves the sparse data problem prevalent in corpus. It adds a new dimension to existing definition of information content as it is defined independent of the corpus statistics. Apart from this it also defines a generalized way of quantifying information content of the concepts, which enables to capture the semantics of the concept. Accordingly, the New Information Content (NIC) based similarity measures NICResnik, NICLin and NICJandC were defined and used to measure the similarity among concepts belonging to the lexical ontology WordNet. The effectiveness of the proposed similarity measures was evaluated using the Psycholinguistic approach. The literature on information retrieval system reveals that similarity measures play a major role in computing document and query similarity. In general, the keywords of the documents are used in the indexing process to retrieve the documents.