Inducing Ontological Co-occurrence Vectors

Proceedings of the 43rd Annual Meeting of the ACL, pages 125–132, Ann Arbor, June 2005. c⃝2005 Association for Computational Linguistics Inducing Ontological Co-occurrence Vectors Patrick Pantel Information Sciences Institute University of Southern California 4676 Admiralty Way Marina del Rey, CA 90292 pantel@isi.edu Abstract In this paper, we present an unsupervised methodology for propagating lexical co- occurrence vectors into an ontology such as WordNet. We evaluate the framework on the task of automatically attaching new concepts into the ontology. Experimental results show 73.9% attachment accuracy in the first position and 81.3% accuracy in the top-5 positions. This framework could potentially serve as a foundation for on- tologizing lexical-semantic resources and assist the development of other large- scale and internally consistent collections of semantic information. 1 Introduction Despite considerable effort, there is still today no commonly accepted semantic corpus, semantic framework, notation, or even agreement on pre- cisely which aspects of semantics are most useful (if at all). We believe that one important reason for this rather startling fact is the absence of truly wide-coverage semantic resources. Recognizing this, some recent work on wide coverage term banks, like WordNet (Miller 1990) and CYC (Lenat 1995), and annotated corpora, like FrameNet (Baker et al. 1998), Propbank (Kingsbury et al. 2002) and Nombank (Meyers et al. 2004), seeks to address the problem. But man- ual efforts such as these suffer from two draw- backs: they are difficult to tailor to new domains, and they have internal inconsistencies that can make automating the acquisition process difficult. In this work, we introduce a general frame- work for inducing co-occurrence feature vectors for nodes in a WordNet-like ontology. We be- lieve that this framework will be useful for a va- riety of applications, including adding additional semantic information to existing semantic term banks by disambiguating lexical-semantic re- sources. Ontologizing semantic resources Recently, researchers have applied text- and web-mining algorithms for automatically creating lexical semantic resources like similarity lists (Lin 1998), semantic lexicons (Riloff and Shep- herd 1997), hyponymy lists (Shinzato and Tori- sawa 2004; Pantel and Ravichandran 2004), part- whole lists (Girgu et al. 2003), and verb relation graphs (Chklovski and Pantel 2004). However, none of these resources have been directly linked into an ontological framework. For example, in VERBOCEAN (Chklovski and Pantel 2004), we find the verb relation “to surpass is-stronger-than to hit”, but it is not specified that it is the achiev- ing sense of hit where this relation applies. We term ontologizing a lexical-semantic re- source as the task of sense disambiguating the re- source. This problem is different but not orthogonal to word-sense disambiguation. If we could disambiguate large collections of text with high accuracy, then current methods for building lexical-semantic resources could easily be applied to ontologize them by treating each word’s senses as separate words. Our method does not require the disambiguation of text. Instead, it relies on the principle of distributional similarity and that polysemous words that are similar in one sense are dissimilar in their other senses. 125 Given …

Content truncated to 500 words