A New Minimally-Supervised Framework for Domain Word Sense Disambiguation

Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pages 1411–1422, Jeju Island, Korea, 12–14 July 2012. c⃝2012 Association for Computational Linguistics A New Minimally-Supervised Framework for Domain Word Sense Disambiguation Stefano Faralli and Roberto Navigli Dipartimento di Informatica Sapienza Universit`a di Roma {faralli,navigli}@di.uniroma1.it Abstract We present a new minimally-supervised framework for performing domain-driven Word Sense Disambiguation (WSD). Glos- saries for several domains are iteratively ac- quired from the Web by means of a boot- strapping technique. The acquired glosses are then used as the sense inventory for fully- unsupervised domain WSD. Our experiments, on new and gold-standard datasets, show that our wide-coverage framework enables high- performance results on dozens of domains at a coarse and fine-grained level. 1 Introduction Domain information pervades most of the text we read every day. If we just think of the Web, the vast majority of its textual content is domain oriented. A case in point is Wikipedia, which provides ency- clopedic coverage for a huge number of knowledge domains (Medelyan et al., 2009), but most blogs, Web sites and newspapers also provide a great deal of information focused on specific areas of knowl- edge. When it comes to automatic text understand- ing, then, it is crucial to take into account the domain specificity of a piece of text, so as to perform a fo- cused and as-precise-as-possible analysis which, in its turn, can enable domain-aware applications such as question answering and information extraction. Domain knowledge also has the potential to improve open-text applications such as summarization (Cey- lan et al., 2010) and machine translation (Foster et al., 2010). Research in Word Sense Disambiguation (Nav- igli, 2009, WSD), the task aimed at the automatic labeling of text with word senses, has been ori- ented towards domain text understanding for sev- eral years now. Many approaches have been devised, including the identification of domain-specific pre- dominant senses (McCarthy et al., 2007; Lapata and Keller, 2007), the development of domain resources (Magnini and Cavagli`a, 2000; Magnini et al., 2002), their application to WSD (Gliozzo et al., 2004), and the effective use of link analysis algorithms such as Personalized PageRank (Agirre et al., 2009; Nav- igli et al., 2011). More recently, semi-supervised ap- proaches to domain WSD have been proposed which aim at decreasing the amount of supervision needed to carry out the task (Khapra et al., 2010). High-performance domain WSD, however, has been hampered by the widespread use of a general- purpose sense inventory, i.e., WordNet (Miller et al., 1990; Fellbaum, 1998). Unfortunately WordNet does not contain many specialized terms, making it difficult to use it in work on arbitrary special- ized domains. While Wikipedia has recently been considered a valid alternative (Mihalcea, 2007), it is mainly focused on covering named entities and, strictly speaking, does not contain a formal wide- coverage sense inventory (not even in disambigua- tion pages, which are often incomplete, especially in the lexicographic sense). In this paper we provide three main contributions: …

Content truncated to 500 words