Evaluating a Semantic Network Automatically Constructed from Lexical Co-occurrence on a Word Sense Disambiguation Task

Proceedings of the Fifteenth Conference on Computational Natural Language Learning, pages 190–199, Portland, Oregon, USA, 23–24 June 2011. c⃝2011 Association for Computational Linguistics Evaluating a Semantic Network Automatically Constructed from Lexical Co-occurrence on a Word Sense Disambiguation Task Sean Szumlanski Department of EECS University of Central Florida seansz@cs.ucf.edu Fernando Gomez Department of EECS University of Central Florida gomez@eecs.ucf.edu Abstract We describe the extension and objective eval- uation of a network1 of semantically related noun senses (or concepts) that has been au- tomatically acquired by analyzing lexical co- occurrence in Wikipedia. The acquisition pro- cess makes no use of the metadata or links that have been manually built into the ency- clopedia, and nouns in the network are auto- matically disambiguated to their correspond- ing noun senses without supervision. For this task, we use the noun sense inventory of WordNet 3.0. Thus, this work can be con- ceived of as augmenting the WordNet noun ontology with unweighted, undirected related- to edges between synsets. Our network con- tains 208,832 such edges. We evaluate our network’s performance on a word sense disambiguation (WSD) task and show: a) the network is competitive with WordNet when used as a stand-alone knowl- edge source for two WSD algorithms; b) com- bining our network with WordNet achieves disambiguation results that exceed the perfor- mance of either resource individually; and c) our network outperforms a similar resource that has been automatically derived from se- mantic annotations in the Wikipedia corpus. 1 Introduction A growing interest in using semantic relatedness in word sense disambiguation (WSD) tasks has spurred investigations into the limitations of the WordNet ontology (Fellbaum, 1998) for this purpose. Al- though WordNet comprises a rich set of semantic 1http://www.cs.ucf.edu/˜seansz/sem links between word senses (or concepts), indicat- ing semantic similarity through subsumptive hyper- nymic and hyponymic relations (among others), it lacks a general indication of semantic relatedness. We present a semantic network that is automat- ically acquired from lexical co-occurrence in Wi- kipedia, and indicates general semantic relatedness between noun senses in WordNet 3.0. In this work, the discovery of relatedness is a context-sparse affair that takes place in absentia of the semantic annota- tions of Wikipedia, such as inter-article links, entries in disambiguation pages, the title of the article from which a sentence is extracted, and so on. We released an earlier version of such a network that was limited by the fact that only relationships involving at least one monosemous noun had been included, and it was not evaluated on a WSD task (Szumlanski and Gomez, 2010). In contrast, the network we present here has relat- edness data for over 4,500 polysemous noun targets and 3,000 monosemous noun targets, each of which are related to an average of 27.5 distinct noun senses. It consists of 208,832 undirected edges – a 181% in- crease in size over the previous network. The result is a semantic network that has reached maturity and, as we will show, can be successfully applied to a WSD task. This paper proceeds as …

Content truncated to 500 words