HARPY: Hypernyms and Alignment of Relational Paraphrases

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 2195–2204, Dublin, Ireland, August 23-29 2014. HARPY: Hypernyms and Alignment of Relational Paraphrases Adam Grycner Max-Planck Institute for Informatics Campus E1.4, 66123 Saarbr¨ucken, Germany agrycner@mpi-inf.mpg.de Gerhard Weikum Max-Planck Institute for Informatics Campus E1.4, 66123 Saarbr¨ucken, Germany weikum@mpi-inf.mpg.de Abstract Collections of relational paraphrases have been automatically constructed from large text cor- pora, as a WordNet counterpart for the realm of binary predicates and their surface forms. How- ever, these resources fall short in their coverage of hypernymy links (subsumptions) among the synsets of phrases. This paper closes this gap by computing a high-quality alignment between the relational phrases of the Patty taxonomy, one of the largest collections of this kind, and the verb senses of WordNet. To this end, we devise judicious features and develop a graph-based alignment algorithm by adapting and extending the SimRank random-walk method. The re- sulting taxonomy of relational phrases and verb senses, coined HARPY, contains 20,812 synsets organized into a Directed Acyclic Graph (DAG) with 616,792 hypernymy links. Our empirical as- sessment, indicates that the alignment links between Patty and WordNet have high accuracy, with Mean Reciprocal Rank (MRR) score 0.7 and Normalized Discounted Cumulative Gain (NDCG) score 0.73. As an additional extrinsic value, HARPY provides fine-grained lexical types for the arguments of verb senses in WordNet. 1 Introduction Motivation: This paper addresses the task of discovering and organizing paraphrases of relations be- tween entities (Lin and Pantel, 2001; Fader et al., 2011; Nakashole et al., 2012; Moro and Navigli, 2012; Alfonseca et al., 2013). This task involves understanding that the phrases “travels to”, “visits” and “on her tour through” (relating a person and a country) are synonymous and that “leader of” and “works with” (relating a person and an organization) are in a hypernymy relation: the former is subsumed by the latter. This kind of lexical knowledge can be harnessed for advanced tasks like question answering (Fader et al., 2013), search over web tables (Gupta et al., 2014), or event mining over news (Alfonseca et al., 2013). Work along these lines has developed large repositories of relational paraphrases, most notably, the collections ReVerb (Fader et al., 2011), Patty (Nakashole et al., 2012), and WiSeNet (Moro and Navigli, 2012). The largest of these, Patty, contains ca. 350,000 synsets of phrases, each annotated with ontolog- ical types of their two arguments (e.g., person × country, or politician × political party). However, the subsumption hierarchy of Patty is very sparse. It contains only 8,000 hypernymy links between phrases, and the entire taxonomy is kind of fragmented into a many-rooted DAG (directed acyclic graph). More- over, the synsets are rather noisy in the long tail with low confidence. WiSeNet, an alternative resource, has ca. 40,000 synsets and no hypernymy links. WordNet (Fellbaum, 1998), on the other hand, is a very rich resource on synonymy and hypernymy. However, its coverage of binary relations (as opposed to unary predicates, mostly nouns) is restricted to (mostly) single-word verbs. …

Content truncated to 500 words