DAEBAK!: Peripheral Diversity for Multilingual Word Sense Disambiguation

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 2: Seventh International Workshop on Semantic Evaluation (SemEval 2013), pages 250–254, Atlanta, Georgia, June 14-15, 2013. c⃝2013 Association for Computational Linguistics DAEBAK!: Peripheral Diversity for Multilingual Word Sense Disambiguation Steve L. Manion University of Canterbury Christchurch, New Zealand steve.manion @pg.canterbury.ac.nz Raazesh Sainudiin University of Canterbury Christchurch, New Zealand r.sainudiin @math.canterbury.ac.nz Abstract We introduce Peripheral Diversity (PD) as a knowledge-based approach to achieve multi- lingual Word Sense Disambiguation (WSD). PD exploits the frequency and diverse use of word senses in semantic subgraphs de- rived from larger sense inventories such as BabelNet, Wikipedia, and WordNet in order to achieve WSD. PD’s f-measure scores for SemEval 2013 Task 12 outperform the Most Frequent Sense (MFS) baseline for two of the five languages: English, French, German, Italian, and Spanish. Despite PD remain- ing under-developed and under-explored, it demonstrates that it is robust, competitive, and encourages development. 1 Introduction By reading out aloud “A minute is a minute divi- sion of time” (Nelson, 1976), we can easily make the distinction between the two senses of the homo- graph minute. For a machine this is a complex task known as Word Sense Disambiguation (WSD). Task 12 of SemEval 2013 (Navigli et al., 2013) calls for a language-independent solution to WSD that utilises a multilingual sense inventory. Supervised approaches to WSD have dominated for some time now (M`arquez et al., 2007). Homo- graphs such as minute are effortlessly disambiguated and more polysemous words such as bar or line can also be disambiguated with reasonable compe- tence (Agirre and Edmonds, 2007). However our ap- proach is purely knowledge-based and employs se- mantic graphs. This allows us to avoid the notorious predicament Gale et al. (1992) name the information bottleneck, in which supervised approaches fail to be portable across alternative languages and domains if the annotated corpora do not exist. Conversely, knowledge-based approaches for WSD are usually applicable to all words in unrestricted text (Mihal- cea, 2007). It is this innate scalability that moti- vates us to pursue knowledge-based approaches. Re- gardless of whether sense inventories can maintain knowledge-richness as they grow, their continued re- finement by contributors is directly beneficial. Knowledge-based approaches that employ se- mantic graphs increasingly rival leading supervised approaches to WSD. They can beat a Random or LESK (Lesk, 1986) baseline (see Mihalcea (2005), Navigli and Lapata (2007), Sinha and Mihalcea (2007), Navigli and Lapata (2010)) and can com- pete with or even beat the Most Frequent Sense (MFS) baseline in certain contexts which is by no means an easy task (see Navigli et al. (2007), Eneko Agirre and Aitor Soroa (2009), Navigli and Ponzetto (2012a)). 2 Methodology PD is a framework for knowledge-based WSD ap- proaches that employ semantic graphs. However be- fore we can elaborate we must first cover the funda- mental resources it is built upon. 2.1 Fundamental Resource Definitions 2.1.1 Lemma Sequences At a glance across the text of any language, we ab- sorb meaning and new information through its lexi- …

Content truncated to 500 words