Million-scale Derivation of Semantic Relations from a Manually Constructed Predicate Taxonomy

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 1423–1434, Dublin, Ireland, August 23-29 2014. Million-scale Derivation of Semantic Relations from a Manually Constructed Predicate Taxonomy Motoki Sano∗ Kentaro Torisawa† Julien Kloetzer‡ Chikara Hashimoto § Istv´an Varga¶ Jong-Hoon Oh∥ ∗ † ‡ § ∥ National Institute of Information and Communications Technology, Kyoto, 619-0289, Japan ¶NEC Knowledge Discovery Research Laboratories, Kanagawa, 211-8666, Japan {∗ msano, † torisawa, ‡ julien, § ch, ∥ rovellia}@nict.go.jp, ¶ vistvan@az.jp.nec.com Abstract We manually created a semantic taxonomy called Phased Predicate Template Taxonomy (PPTT) that covers 12,023 predicate templates (i.e., predicates with one argument slot like “rescue X”) and derived from it various semantic relations between these templates on a million-instance scale (70%-80% precision level). The derived relations include entailment (e.g., rescue X⊃X is alive), happens-before (e.g., buy X⇒drink X), and a novel relation type anomalous obstruction (e.g., X is sold out;cannot buy X). Such derivation became possible thanks to PPTT’s design and the use of statistical methods. 1 Introduction Databases of various semantic relations between natural language expressions are indispensable knowl- edge for many NLP applications. For instance, entailment relations are crucial in information extraction and QA (Dagan et al., 2009; Weisman et al., 2012; Berant et al., 2012; Turney and Mohammad, 2014). Temporal relations such as happens-before (Chklovski and Pantel, 2004b; Regneri et al., 2010) are im- portant for enhancing deep semantic processing. A problem, however, is that it is difficult to acquire those relations with a broad coverage. Although many sophisticated machine learning techniques have been applied to various kinds of corpora for this task (Szpektor et al., 2007; Chambers and Jurafsky, 2008; Hashimoto et al., 2009; Chambers and Jurafsky, 2009; Hashimoto et al., 2012; Talukdar et al., 2012; Kloetzer et al., 2013), no satisfactory coverage has been achieved, probably due to data sparseness in the input data. In this work we take a completely different approach: we manually construct a seman- tic lexicon called Phased Predicate Template Taxonomy (PPTT), and derive various types of semantic relations on a large-scale by using it. Our target language is Japanese, but examples are given in English for simplicity throughout this paper. PPTT is a taxonomy of predicate templates (predicates with one argument slot like rescue X, “Tem- plate” hereafter) that classifies templates according to phases of story concerning an entity denoted by X. In the story, or the “life” of the entity X, X can be anticipated, created, then execute its function and finally it may collapse and become deficient. Anticipation, creation, execution, collapse, deficiency of X can be seen as such phases of story concerning X, and PPTT classifies templates into 41 semantic classes each of which corresponds to a phase. In other words, PPTT provides a way to describe the stories of var- ious entities that constitute this world, and we believe that PPTT (partly) reflects how we understand the world and its entities. Accordingly, PPTT can also provide a way to derive various semantic knowledge about this world …

Content truncated to 500 words