SPred: Large-scale Harvesting of Semantic Predicates

Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics, pages 1222–1232, Sofia, Bulgaria, August 4-9 2013. c⃝2013 Association for Computational Linguistics SPred: Large-scale Harvesting of Semantic Predicates Tiziano Flati and Roberto Navigli Dipartimento di Informatica Sapienza Universit`a di Roma {flati,navigli}@di.uniroma1.it Abstract We present SPred, a novel method for the creation of large repositories of semantic predicates. We start from existing colloca- tions to form lexical predicates (e.g., break ∗) and learn the semantic classes that best fit the ∗ argument. To do this, we extract all the occurrences in Wikipedia which match the predicate and abstract its argu- ments to general semantic classes (e.g., break BODY PART, break AGREEMENT, etc.). Our experiments show that we are able to create a large collection of seman- tic predicates from the Oxford Advanced Learner’s Dictionary with high precision and recall, and perform well against the most similar approach. 1 Introduction Acquiring semantic knowledge from text automat- ically is a long-standing issue in Computational Linguistics and Artificial Intelligence. Over the last decade or so the enormous abundance of in- formation and data that has become available has made it possible to extract huge amounts of pat- terns and named entities (Etzioni et al., 2005), se- mantic lexicons for categories of interest (Thelen and Riloff, 2002; Igo and Riloff, 2009), large do- main glossaries (De Benedictis et al., 2013) and lists of concepts (Katz et al., 2003). Recently, the availability of Wikipedia and other collabora- tive resources has considerably boosted research on several aspects of knowledge acquisition (Hovy et al., 2013), leading to the creation of several large-scale knowledge resources, such as DBPe- dia (Bizer et al., 2009), BabelNet (Navigli and Ponzetto, 2012), YAGO (Hoffart et al., 2013), MENTA (de Melo and Weikum, 2010), to name but a few. This wealth of acquired knowledge is known to have a positive impact on important fields such as Information Retrieval (Chu-Carroll and Prager, 2007), Information Extraction (Krause et al., 2012), Question Answering (Ferrucci et al., 2010) and Textual Entailment (Berant et al., 2012; Stern and Dagan, 2012). Not only are these knowledge resources ob- tained by acquiring concepts and named entities, but they also provide semantic relations between them. These relations are extracted from unstruc- tured or semi-structured text using ontology learn- ing from scratch (Velardi et al., 2013) and Open Information Extraction techniques (Etzioni et al., 2005; Yates et al., 2007; Wu and Weld, 2010; Fader et al., 2011; Moro and Navigli, 2013) which mainly stem from seminal work on is-a relation acquisition (Hearst, 1992) and subsequent devel- opments (Girju et al., 2003; Pasca, 2004; Snow et al., 2004, among others). However, these knowledge resources still lack semantic information about language units such as phrases and collocations. For instance, which semantic classes are expected as a direct object of the verb break? What kinds of noun does the adjective amazing collocate with? Recognition of the need for systems that are aware of the selec- tional restrictions of verbs and, more in general, of …

Content truncated to 500 words