Extending WordNet with Fine-Grained Collocational Information via Supervised Distributional Learning

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 3422–3432, Osaka, Japan, December 11-17 2016. Extending WordNet with Fine-Grained Collocational Information via Supervised Distributional Learning Luis Espinosa-Anke1, Jose Camacho-Collados2, Sara Rodr´ıguez-Fern´andez1, Horacio Saggion1, Leo Wanner3,1 1NLP Group, Universitat Pompeu Fabra firstname.lastname@upf.edu 2Department of Computer Science, Sapienza University of Rome collados@di.uniroma1.it 3Catalan Institute for Research and Advanced Studies (ICREA) Abstract WordNet is probably the best known lexical resource in Natural Language Processing. While it is widely regarded as a high quality repository of concepts and semantic relations, updating and extending it manually is costly. One important type of relation which could potentially add enor- mous value to WordNet is the inclusion of collocational information, which is paramount in tasks such as Machine Translation, Natural Language Generation and Second Language Learning. In this paper, we present ColWordNet (CWN), an extended WordNet version with fine-grained col- locational information, automatically introduced thanks to a method exploiting linear relations between analogous sense-level embeddings spaces. We perform both intrinsic and extrinsic eval- uations, and release CWN for the use and scrutiny of the community. 1 Introduction The embedding of cues about how we perceive concepts and how these concepts relate and generalize across different domains gives knowledge resources the capacity of generalization, which lies at the core of human cognition (Yu et al., 2015) and is also central to many Natural Language Processing (NLP) applications (Jurgens and Pilehvar, 2015). It is general practice to identify and formalize conceptual relations using a reference knowledge repository. As such a repository, WordNet (Miller et al., 1990) stands out as the de facto standard lexical database, containing over 200k English senses with 155k word forms. Over the years, WordNet has become the cornerstone of agglutinative resources such as BabelNet (Navigli and Ponzetto, 2012) and Yago (Suchanek et al., 2007). It is also used in semantically intensive tasks such as Word Sense Disambiguation (Navigli, 2009), Query Expansion and IR (Fang, 2008), Sentiment Analysis (Esuli and Sebastiani, 2006), semantic similarity measurement (Pilehvar et al., 2013), development and evaluation of word embeddings models (Huang et al., 2012; Faruqui et al., 2015), and Taxonomy Learning Evaluation (Bordea et al., 2015). While the value of WordNet for NLP is indisputable, it is generally recognized that enriching it with additional information makes it an even more valuable resource. Thus, there is a line of research aimed at extending it with novel terminology (Jurgens and Pilehvar, 2016), cross-predicate relations (Lopez de la Calle et al., 2016), and so forth. Nonetheless, one type of information has been largely neglected so far: collocations, i.e., idiosyncratic binary lexical co-occurrences. As a standalone research topic, however, collocations have been the focus of a substantial amount of work, e.g. for automatically retrieving them from corpora (Choueka, 1988; Church and Hanks, 1989; Smadja, 1993; Kilgariff, 2006; Evert, 2007; Pecina, 2008; Bouma, 2010; Gao, 2013), and for their semantic classification according to different typologies (Wanner et al., 2006; Gelbukh and Kolesnikova., 2012; Moreno et al., 2013; Wanner et al., …

Content truncated to 500 words