A Study on the Semantic Relatedness of Query and Document Terms in Information Retrieval

Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 1338–1347, Singapore, 6-7 August 2009. c⃝2009 ACL and AFNLP A Study on the Semantic Relatedness of Query and Document Terms in Information Retrieval Christof M¨uller and Iryna Gurevych Ubiquitous Knowledge Processing (UKP) Lab Computer Science Department Technische Universit¨at Darmstadt, Hochschulstraße 10 D-64289 Darmstadt, Germany http://www.ukp.tu-darmstadt.de/ Abstract The use of lexical semantic knowledge in information retrieval has been a field of ac- tive study for a long time. Collaborative knowledge bases like Wikipedia and Wik- tionary, which have been applied in com- putational methods only recently, offer new possibilities to enhance information retrieval. In order to find the most bene- ficial way to employ these resources, we analyze the lexical semantic relations that hold among query and document terms and compare how these relations are repre- sented by a measure for semantic related- ness. We explore the potential of different indicators of document relevance that are based on semantic relatedness and com- pare the characteristics and performance of the knowledge bases Wikipedia, Wik- tionary and WordNet. 1 Introduction Today we face a rapidly growing number of elec- tronic documents in all areas of life. This demands for more effective and efficient ways of searching these documents for information. Especially user- generated content on the web is a growing source of huge amounts of data that poses special diffi- culties to IR. The precise wording is often difficult to predict and current information retrieval (IR) systems are mainly based on the assumption that the meaning of a document can be inferred from the occurrence or absence of terms in it. In or- der to yield a good retrieval performance, i.e., re- trieving all relevant documents without retrieving non-relevant documents, the query has to be for- mulated by the user in an appropriate way. Blair and Maron (1985) showed that with larger grow- ing document collections, it gets impossible for the user to anticipate the terms that occur in all relevant documents, but not in non-relevant ones. The use of semantic knowledge for improving IR by compensating non-optimal queries has been a field of study for a long time. First experi- ments on query expansion by Voorhees (1994) us- ing lexical-semantic relations extracted from a lin- guistic knowledge base (LKB), namely WordNet (Fellbaum, 1998), showed sginificant improve- ments in performance only for manually selected expansion terms. The combination of Word- Net with thesauri built from the underlying doc- ument collections by Mandala et al. (1998) im- proved the performance on several test collec- tions. Mandala et al. (1998) identified missing relations, especially cross part of speech relations and insufficient lexical coverage as reasons for the low performance improvement when using only WordNet. In recent work, collaborative knowledge bases (CKB) like Wikipedia have been used in IR for judging the document relevance by computing the semantic relatedness (SR) of queries and docu- ments (Gurevych et al., 2007; Egozi et al., 2008; M¨uller and Gurevych, 2008) and have shown promising results. These …

Content truncated to 500 words