Learning Arguments and Supertypes of Semantic Relations Using Recursive Patterns

Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 1482–1491, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Learning Arguments and Supertypes of Semantic Relations using Recursive Patterns Zornitsa Kozareva and Eduard Hovy USC Information Sciences Institute 4676 Admiralty Way Marina del Rey, CA 90292-6695 {kozareva,hovy}@isi.edu Abstract A challenging problem in open informa- tion extraction and text mining is the learn- ing of the selectional restrictions of se- mantic relations. We propose a mini- mally supervised bootstrapping algorithm that uses a single seed and a recursive lexico-syntactic pattern to learn the ar- guments and the supertypes of a diverse set of semantic relations from the Web. We evaluate the performance of our algo- rithm on multiple semantic relations ex- pressed using “verb”, “noun”, and “verb prep” lexico-syntactic patterns. Human- based evaluation shows that the accuracy of the harvested information is about 90%. We also compare our results with existing knowledge base to outline the similarities and differences of the granularity and di- versity of the harvested knowledge. 1 Introduction Building and maintaining knowledge-rich re- sources is of great importance to information ex- traction, question answering, and textual entail- ment. Given the endless amount of data we have at our disposal, many efforts have focused on mining knowledge from structured or unstructured text, including ground facts (Etzioni et al., 2005), se- mantic lexicons (Thelen and Riloff, 2002), ency- clopedic knowledge (Suchanek et al., 2007), and concept lists (Katz et al., 2003). Researchers have also successfully harvested relations between en- tities, such as is-a (Hearst, 1992; Pasca, 2004) and part-of (Girju et al., 2003). The kinds of knowl- edge learned are generally of two kinds: ground instance facts (New York is-a city, Rome is the cap- ital of Italy) and general relational types (city is-a location, engines are part-of cars). A variety of NLP tasks involving inference or entailment (Zanzotto et al., 2006), including QA (Katz and Lin, 2003) and MT (Mt et al., 1988), require a slightly different form of knowledge, de- rived from many more relations. This knowledge is usually used to support inference and is ex- pressed as selectional restrictions (Wilks, 1975) (namely, the types of arguments that may fill a given relation, such as person live-in city and air- line fly-to location). Selectional restrictions con- strain the possible fillers of a relation, and hence the possible contexts in which the patterns ex- pressing that relation can participate in, thereby enabling sense disambiguation of both the fillers and the expression itself. To acquire this knowledge two common ap- proaches are employed: clustering and patterns. While clustering has the advantage of being fully unsupervised, it may or may not produce the types and granularity desired by a user. In contrast pattern-based approaches are more precise, but they typically require a handful to dozens of seeds and lexico-syntactic patterns to initiate the learn- ing process. In a closed domain these approaches are both very promising, but when tackling an un- bounded number of relations they …

Content truncated to 500 words