UoS: A Graph-Based System for Graded Word Sense Induction

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 2: Seventh International Workshop on Semantic Evaluation (SemEval 2013), pages 689–694, Atlanta, Georgia, June 14-15, 2013. c⃝2013 Association for Computational Linguistics UoS: A Graph-Based System for Graded Word Sense Induction David Hope, Bill Keller University of Sussex Cognitive and Language Processing Systems Group Brighton, Sussex, UK davehope@gmail.com, billk@sussex.ac.uk Abstract This paper presents UoS, a graph-based Word Sense Induction system which attempts to find all applicable senses of a target word given its context, grading each sense accord- ing to its suitability to the context. Senses of a target word are induced through use of a non-parameterised, linear-time clustering al- gorithm that returns maximal quasi-strongly connected components of a target word graph in which vertex pairs are assigned to the same cluster if either vertex has the highest edge weight to the other. UoS participated in SemEval-2013 Task 13: Word Sense Induc- tion for Graded and Non-Graded Senses. Two system were submitted; both systems returned results comparable with those of the best per- forming systems. 1 Introduction Word Sense Induction (WSI) is the task of automat- ically discovering word senses from text. In princi- ple, WSI avoids reliance on a pre-defined sense in- ventory.1 Whereas the related task of Word Sense Disambiguation (WSD) can only assign pre-defined senses to words on the basis of context, WSI fol- lows the dictum that “The meaning of a word is its use in the language.” (Wittgenstein, 1953) to dis- cover senses through examination of context of use in large text corpora. WSI, therefore, may be applied 1In practice, evaluation of a WSI system requires the use of a gold standard sense inventory such as WordNet (Miller et al., 1990) or OntoNotes (Hovy et al., 2006). to discover new, rare, or domain specific senses; senses undefined in existing sense inventories.2 Previous WSI evaluations (Agirre and Soroa, 2007; Manandhar et al., 2010) have approached sense induction in terms of finding the single most salient sense of a target word given its context. However, as shown in Erk and McCarthy (2009), a graded notion of sense may be more applicable, as multiple senses of the target word may be perceived by readers. The SemEval-2013 WSI evaluation de- scribed in this paper is designed to explore the possi- bility of finding all perceived senses of a target word in a single contextual instance. The aim for partici- pants in the task is therefore to design a system that will induce a set of graded (weighted) senses of a target word in a particular context. The paper is organised as follows: Section 2 in- troduces SemEval-2013 Task 13: Word Sense In- duction for Graded and Non-Graded Senses; Sec- tion 3 presents UoS, the system that participated in the task; Section 4 reports evaluation results, show- ing that UoS returns scores comparable with those of the best performing systems. 2 SemEval-2013 Task 13 2.1 Aim The aim for participants in SemEval-2013 Task 13: Word Sense Induction for Graded and Non-Graded …

Content truncated to 500 words