unimelb: Topic Modelling-based Word Sense Induction for Web Snippet Clustering

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 2: Seventh International Workshop on Semantic Evaluation (SemEval 2013), pages 217–221, Atlanta, Georgia, June 14-15, 2013. c⃝2013 Association for Computational Linguistics unimelb: Topic Modelling-based Word Sense Induction for Web Snippet Clustering Jey Han Lau, Paul Cook and Timothy Baldwin Department of Computing and Information Systems The University of Melbourne jhlau@csse.unimelb.edu.au, paulcook@unimelb.edu.au, tb@ldwin.net Abstract This paper describes our system for Task 11 of SemEval-2013. In the task, participants are provided with a set of ambiguous search queries and the snippets returned by a search engine, and are asked to associate senses with the snippets. The snippets are then clus- tered using the sense assignments and sys- tems are evaluated based on the quality of the snippet clusters. Our system adopts a pre- existing Word Sense Induction (WSI) method- ology based on Hierarchical Dirichlet Process (HDP), a non-parametric topic model. Our system is trained over extracts from the full text of English Wikipedia, and is shown to per- form well in the shared task. 1 Introduction The basic premise behind research on word sense disambiguation (WSD) is that there exists a static, discrete set of word senses that can be used to la- bel distinct usages of a given word (Agirre and Ed- monds, 2006; Navigli, 2009). There are various pit- falls underlying this premise, including: (1) what sense inventory is appropriate for a particular task (given that sense inventories can vary considerably in their granularity and partitioning of word usages)? (2) given that word senses tend to take the form of prototypes, is discrete labelling a felicitous represen- tation of word usages, especially for non-standard word usages? (3) how should novel word usages be captured under this model? and (4) given the rapid pace of language evolution on real-time social me- dia such as Twitter and Facebook, is it reasonable to assume a static sense inventory? Given this back- drop, there has been a recent growth of interest in the task of word sense induction (WSI), where the word sense representation for a given word is automati- cally inferred from a given data source, and word usages are labelled (often probabilistically) accord- ing to that data source. While WSI has considerable appeal as a task, intrinsic cross-comparison of WSI systems is fraught with many of the same issues as WSD (Agirre and Soroa, 2007; Manandhar et al., 2010), leading to a move towards task-based WSI evaluation, such as in Task 11 of SemEval-2013, ti- tled “Evaluating Word Sense Induction & Disam- biguation within an End-User Application”. This paper presents the UNIMELB system entry to SemEval-2013 Task 11. Our method is based heav- ily on the WSI methodology proposed by Lau et al. (2012) for novel word sense detection. Largely the same methodology was also applied to SemEval- 2013 Task 13 on WSI (Lau et al., to appear). 2 System Description Our system is based on the WSI methodology pro- posed by Lau et al. (2012) for the task of novel word …

Content truncated to 500 words