Harnessing different knowledge sources to measure semantic relatedness under a uniform model

Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing, pages 991–1002, Edinburgh, Scotland, UK, July 27–31, 2011. c⃝2011 Association for Computational Linguistics Harnessing different knowledge sources to measure semantic relatedness under a uniform model Ziqi Zhang Anna Lisa Gentile Fabio Ciravegna Department of Computer Science, University of Sheffield 211 Portobello, Regent Court Sheffield, S1 4DP z.zhang@dcs.shef.ac. uk a.l.gentile@dcs.shef .ac.uk f.ciravegna@dcs.shef .ac.uk Abstract Measuring semantic relatedness between words or concepts is a crucial process to many Natural Language Processing tasks. Exiting methods exploit semantic evidence from a single knowledge source, and are predominantly evaluated only in the general domain. This paper introduces a method of harnessing different knowledge sources under a uniform model for measuring semantic relatedness between words or concepts. Using Wikipedia and WordNet as examples, and evaluated in both the general and biomedical domains, it successfully combines strengths from both knowledge sources and outperforms state- of-the-art on many datasets. 1 Introduction Semantic relatedness (SR) measures how much two (strings of) words or concepts are related by encompassing all kinds of relations between them (Strube and Ponzetto, 2006). It is more general than semantic similarity. SR is often an important pre-processing step to many complex Natural Language Processing (NLP) tasks, such as Word Sense Disambiguation (Leacock and Chodorow, 1998; Han and Zhao, 2010), and information retrieval (Finkelstein et al., 2002). In the biomedical domain, SR is an important technique for discovering gene functions and interactions (Wu et al., 2005; Ye et al., 2005). There is an abundant literature on measuring SR between words or concepts. Typically, these methods extract semantic evidence of words and concepts from a background knowledge source, with which their relatedness is assessed. The knowledge sources can be unstructured documents or (semi-)structured resources such as Wikipedia, WordNet, and domain specific ontologies (e.g., the Gene Ontology1). In this paper, we identify two issues that have not been addressed in the previous works. First, existing works typically employ a single knowledge source of semantic evidence. Research (Strube and Ponzetto, 2006; Zesch and Gurevych, 2010; Zhang et al., 2010) has shown that the accuracy of an SR method differs depending on the choice of the knowledge sources, and there is no conclusion which knowledge source is superior to others. Zhang et al. (2010) argue that this indicates different knowledge sources may complement each other. Second, the majority of SR methods have been evaluated in general domains only, except a few earlier WordNet-based methods that have been adapted to biomedical ontologies and evaluated in that domain (Lord et al., 2003; Pedersen et al., 2006; Pozo et al., 2008). Given the significant attention that SR has received in specific domains (Pesquita et al., 2007), evaluation of SR methods in specific domains is increasingly important. This paper addresses these issues by proposing a generic and uniform model for computing SR between words or concepts using multiple knowledge sources, and evaluating the proposed method in both general and specific domains. The method combines and integrates semantic evidence of words or concepts …

Content truncated to 500 words