Refining the Notions of Depth and Density in WordNet-based Semantic Similarity Measures

Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing, pages 1003–1011, Edinburgh, Scotland, UK, July 27–31, 2011. c⃝2011 Association for Computational Linguistics Refining the Notions of Depth and Density in WordNet-based Semantic Similarity Measures Tong Wang Department of Computer Science University of Toronto tong@cs.toronto.edu Graeme Hirst Department of Computer Science University of Toronto gh@cs.toronto.edu Abstract We re-investigate the rationale for and the ef- fectiveness of adopting the notions of depth and density in WordNet-based semantic sim- ilarity measures. We show that the intuition for including these notions in WordNet-based similarity measures does not always stand up to empirical examination. In particular, the traditional definitions of depth and density as ordinal integer values in the hierarchical structure of WordNet does not always corre- late with human judgment of lexical semantic similarity, which imposes strong limitations on their contribution to an accurate similarity measure. We thus propose several novel defi- nitions of depth and density, which yield sig- nificant improvement in degree of correlation with similarity. When used in WordNet-based semantic similarity measures, the new defini- tions consistently improve performance on a task of correlating with human judgment. 1 Introduction Semantic similarity measures are widely used in natural language processing for measuring distance between meanings of words. There are currently two mainstream approaches to deriving such mea- sures, i.e., distributional and lexical resource-based approaches. The former usually explores the co- occurrence patterns of words in large collections of texts such as text corpora (Lin, 1998) or the Web (Turney, 2001). The latter takes advantage of mostly handcrafted information, such as dictionar- ies (Chodorow et al., 1985; Kozima and Ito, 1997) or thesauri (Jarmasz and Szpakowicz, 2003). Another important resource in the latter stream is semantic taxonomies such as WordNet (Fellbaum, 1998). Despite their high cost of compilation and limited availability across languages, semantic tax- onomies have been widely used in similarity mea- sures, and one of the main reasons behind this is that the often complex notion of lexical semantic simi- larity can be approximated with ease by the distance between words (represented as nodes) in their hier- archical structures, and this approximation appeals much to our intuition. Even methods as simple as “hop counts” between nodes (e.g., that of Rada et al. 1989 on the English WordNet) can take us a long way. Meanwhile, taxonomy-based methods have been constantly refined by incorporating various structural features such as depth (Sussna, 1993; Wu and Palmer, 1994), density (Sussna, 1993), type of connection (Hirst and St-Onge, 1998; Sussna, 1993), word class (sense) frequency estimates (Resnik, 1999), or a combination these features (Jiang and Conrath, 1997). Most of these algorithms are fairly self-contained and easy to implement, with off-the- shelf toolkits such as that of Pedersen et al. (2004). With the existing literature focusing on carefully weighting these features to construct a better seman- tic similarity measure, however, the rationale for adopting these features in calculating semantic sim- ilarity remains largely intuitive. To the best of our knowledge, there is no empirical …

Content truncated to 500 words