Unsupervised Domain Tuning to Improve Word Sense Disambiguation

Proceedings of NAACL-HLT 2013, pages 680–684, Atlanta, Georgia, 9–14 June 2013. c⃝2013 Association for Computational Linguistics Unsupervised Domain Tuning to Improve Word Sense Disambiguation Judita Preiss and Mark Stevenson j.preiss@sheffield.ac.uk and m.stevenson@dcs.shef.ac.uk Department of Computer Science, University of Sheffield 211 Portobello, Sheffield, S1 4DP, UK Abstract The topic of a document can prove to be use- ful information for Word Sense Disambigua- tion (WSD) since certain meanings tend to be associated with particular topics. This paper presents an LDA-based approach for WSD, which is trained using any available WSD sys- tem to establish a sense per (Latent Dirich- let allocation based) topic. The technique is tested using three unsupervised and one su- pervised WSD algorithms within the SPORT and FINANCE domains giving a performance increase each time, suggesting that the tech- nique may be useful to improve the perfor- mance of any available WSD system. 1 Introduction Assigning each word its most frequent sense (MFS) is commonly used as a baseline in Word Sense Dis- ambiguation (WSD). This baseline can be difficult to beat, particularly for unsupervised systems which do not have access to the annotated training data used to determine the MFS. However, it has also been shown that unsupervised methods can be used to identify the most likely sense for each ambiguous word type and this approach can be effective for disambigua- tion (McCarthy et al., 2004). Knowledge of the domain of a document has been shown to be useful information for WSD. For ex- ample, Khapra et al. (2010) improve the perfor- mance of a graph-based WSD system using a small number of hand-tagged examples, but further ex- amples would be required for each new domain. Agirre et al. (2009) automatically construct a the- saurus from texts in a domain which they use for WSD. Unfortunately, performance drops when the thesaurus is combined with information from local context. Stevenson et al. (2011) showed that per- formance of an unsupervised WSD algorithm can be improved by supplementing the context with do- main information. Cai et al. (2007) use LDA to create an additional feature for a supervised WSD algorithm, by inferring topics for labeled training data. Boyd-Graber et al. (2007) integrate a topic model with WordNet and use it to carry out dis- ambiguation and learn topics simultaneously. Li et al. (2010) use sense paraphrases to estimate prob- abilities of senses and carry out WSD. Koeling et al. (2005) showed that automatically acquiring the predominant sense of a word from a corpus from the same domain increases performance (over using a predominant sense acquired from a balanced cor- pus), but their work requires a separate thesaurus to be built for each domain under investigation. Nav- igli et al. (2011) extracted relevant terms from texts in a domain and used them to initialize a random walk over the WordNet graph. Our approaches rely on a one sense per topic hypothesis (Gale et al., 1992), making use of top- ics induced using LDA – we present three novel techniques for …

Content truncated to 500 words