Estimating Semantic Distance Using Soft Semantic Constraints in Knowledge-Source – Corpus Hybrid Models

Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 775–783, Singapore, 6-7 August 2009. c⃝2009 ACL and AFNLP Estimating Semantic Distance Using Soft Semantic Constraints in Knowledge-Source–Corpus Hybrid Models Yuval Marton∗†, Saif Mohammad†, and Philip Resnik∗† ∗Department of Linguistics and †Laboratory for Computational Linguistics and Information Processing, Institute for Advanced Computer Studies. University of Maryland, College Park, MD 20742-7505, USA. {ymarton,saif,resnik}@umiacs.umd.edu Abstract Strictly corpus-based measures of seman- tic distance conflate co-occurrence infor- mation pertaining to the many possible senses of target words. We propose a corpus–thesaurus hybrid method that uses soft constraints to generate word-sense- aware distributional profiles (DPs) from coarser “concept DPs” (derived from a Roget-like thesaurus) and sense-unaware traditional word DPs (derived from raw text). Although it uses a knowledge source, the method is not vocabulary- limited: if the target word is not in the thesaurus, the method falls back grace- fully on the word’s co-occurrence infor- mation. This allows the method to access valuable information encoded in a lexical resource, such as a thesaurus, while still being able to effectively handle domain- specific terms and named entities. Exper- iments on word-pair ranking by semantic distance show the new hybrid method to be superior to others. 1 Introduction Semantic distance is a measure of the closeness in meaning of two concepts. People are consis- tent judges of semantic distance. For example, we can easily tell that the concepts of “exercise” and “jog” are closer in meaning than “exercise” and “theater”. Studies asking native speakers of a lan- guage to rank word pairs in order of semantic dis- tance confirm this—average inter-annotator corre- lation on ranking word pairs in order of semantic distance has been repeatedly shown to be around 0.9 (Rubenstein and Goodenough, 1965; Resnik, 1999). A number of natural language tasks such as ma- chine translation (Lopez, 2008) and word sense disambiguation (Banerjee and Pedersen, 2003; McCarthy, 2006), can be framed as semantic distance problems. Thus, developing automatic measures that are in-line with human notions of semantic distance has received much attention. These automatic approaches to semantic distance rely on manually created lexical resources such as WordNet, large amounts of text corpora, or both. WordNet-based information content measures have been successful (Hirst and Budanitsky, 2005), but there are significant limitations on their applicability. They can be applied only if a Word- Net exists for the language of interest (which is not the case for the “low-density” languages); and even if there is a WordNet, a number of domain- specific terms may not be encoded in it. On the other hand, corpus-based distributional measures of semantic distance, such as cosine and α-skew divergence (Dagan et al., 1999), rely on raw text alone (Weeds et al., 2004; Mohammad, 2008). However, when used to rank word pairs in order of semantic distance or correct real-word spelling errors, they have been shown to perform poorly (Weeds et al., 2004; Mohammad and Hirst, 2006). Mohammad and Hirst (2006) and Patwardhan and Pedersen (2006) argued that word sense ambi- …

Content truncated to 500 words