NASARI: a Novel Approach to a Semantically-Aware Representation of Items

Human Language Technologies: The 2015 Annual Conference of the North American Chapter of the ACL, pages 567–577, Denver, Colorado, May 31 – June 5, 2015. c⃝2015 Association for Computational Linguistics NASARI: a Novel Approach to a Semantically-Aware Representation of Items Jos´e Camacho-Collados, Mohammad Taher Pilehvar and Roberto Navigli Department of Computer Science Sapienza University of Rome {collados,pilehvar,navigli}@di.uniroma1.it Abstract The semantic representation of individual word senses and concepts is of fundamental importance to several applications in Natu- ral Language Processing. To date, concept modeling techniques have in the main based their representation either on lexicographic re- sources, such as WordNet, or on encyclope- dic resources, such as Wikipedia. We pro- pose a vector representation technique that combines the complementary knowledge of both these types of resource. Thanks to its use of explicit semantics combined with a novel cluster-based dimensionality reduction and an effective weighting scheme, our repre- sentation attains state-of-the-art performance on multiple datasets in two standard bench- marks: word similarity and sense clustering. We are releasing our vector representations at http://lcl.uniroma1.it/nasari/. 1 Introduction Obtaining accurate semantic representations of indi- vidual word senses or concepts is vital for several applications in Natural Language Processing (NLP) such as, for example, Word Sense Disambigua- tion (Navigli, 2009; Navigli, 2012), Entity Linking (Bunescu and Pas¸ca, 2006; Rao et al., 2013), seman- tic similarity (Budanitsky and Hirst, 2006), Informa- tion Extraction (Banko et al., 2007), and resource linking and integration (Pilehvar and Navigli, 2014). One prominent semantic representation approach is the distributional semantic model, which represents lexical items as vectors in a semantic space. The weights in these vectors were traditionally computed on the basis of co-occurrence statistics (Salton et al., 1975; Turney and Pantel, 2010; Dinu and Lapata, 2010; Lappin and Fox, 2014), whereas for the more recent generation of distributional models weight computation is viewed as a context prediction prob- lem, often to be solved by using neural networks (Collobert and Weston, 2008; Turian et al., 2010; Mikolov et al., 2013). Unfortunately, unless they are provided with large amounts of sense-annotated data these corpus-based techniques cannot capture polysemy in their representations, as they conflate different meanings of a word into a single vector. Therefore, most sense modeling techniques tend to base their computation on the knowledge obtained from various lexical resources. However, these tech- niques mainly utilize the knowledge derived from either WordNet (Banerjee and Pedersen, 2002; Bu- danitsky and Hirst, 2006; Pilehvar et al., 2013) or Wikipedia (Medelyan et al., 2009; Mihalcea, 2007; Dandala et al., 2013; Gabrilovich and Markovitch, 2007; Strube and Ponzetto, 2006), which are, re- spectively, the most widely-used lexicographic and encyclopedic resources in lexical semantics (Hovy et al., 2013). This restriction to a single resource brings about two main limitations: (1) the sense modeling does not benefit from the complementary knowledge of different resources, and (2) the ob- tained representations are resource-specific and can- not be used across settings. In this paper we put forward a novel concept rep- resentation technique, called NASARI, which ex- ploits the knowledge …

Content truncated to 500 words