Supervised Distributional Hypernym Discovery via Domain Adaptation

Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 424–435, Austin, Texas, November 1-5, 2016. c⃝2016 Association for Computational Linguistics Supervised Distributional Hypernym Discovery via Domain Adaptation Luis Espinosa-Anke1, Jose Camacho-Collados2, Claudio Delli Bovi2 and Horacio Saggion1 1Department of Information and Communication Technologies, Universitat Pompeu Fabra 2Department of Computer Science, Sapienza University of Rome 1{luis.espinosa,horacio.saggion}@upf.edu 2{collados,dellibovi}@di.uniroma1.it Abstract Lexical taxonomies are graph-like hierarchi- cal structures that provide a formal representa- tion of knowledge. Most knowledge graphs to date rely on is-a (hypernymic) relations as the backbone of their semantic structure. In this paper, we propose a supervised distributional framework for hypernym discovery which op- erates at the sense level, enabling large-scale automatic acquisition of disambiguated tax- onomies. By exploiting semantic regularities between hyponyms and hypernyms in embed- dings spaces, and integrating a domain clus- tering algorithm, our model becomes sensi- tive to the target data. We evaluate several configurations of our approach, training with information derived from a manually created knowledge base, along with hypernymic rela- tions obtained from Open Information Extrac- tion systems. The integration of both sources of knowledge yields the best overall results ac- cording to both automatic and manual evalua- tion on ten different domains. 1 Introduction Lexical taxonomies (taxonomies henceforth) are graph-like hierarchical structures where terms are nodes, and are typically organized over a predefined merging or splitting criterion (Hwang et al., 2012). By embedding cues about how we perceive con- cepts, and how these concepts generalize in a do- main of knowledge, these resources bear a capacity for generalization that lies at the core of human cog- nition (Yu et al., 2015) and have become key in Nat- ural Language Processing (NLP) tasks where infer- ence and reasoning have proved to be essential. In fact, taxonomies have enabled a remarkable number of novel NLP techniques, e.g. the contribution of WordNet (Miller, 1995) to lexical semantics (Pile- hvar et al., 2013; Yu and Dredze, 2014) as well as various tasks, from word sense disambiguation (Agirre et al., 2014) to information retrieval (Vare- las et al., 2005), question answering (Harabagiu et al., 2003) and textual entailment (Glickman et al., 2005). To date, the application of taxonomies in NLP has consisted mainly of, on one hand, formally representing a domain of knowledge (e.g. Food), and, on the other hand, constituting the semantic backbone of large-scale knowledge repositories such as ontologies or Knowledge Bases (KBs). In domain knowledge formalization, prominent work has made use of the web (Kozareva and Hovy, 2010), lexico-syntactic patterns (Navigli and Ve- lardi, 2010), syntactic evidence (Luu Anh et al., 2014), graph-based algorithms (Fountain and Lap- ata, 2012; Velardi et al., 2013; Bansal et al., 2014) or popularity of web sources (Luu Anh et al., 2015). As for enabling large-scale knowledge repositories, this task often tackles the additional problem of disam- biguating word senses and entity mentions. Notable approaches of this kind include Yago (Suchanek et al., 2007), WikiTaxonomy (Ponzetto and Strube, 2008), and the Wikipedia Bitaxonomy (Flati et al., 2014). …

Content truncated to 500 words