Automatic Discovery of Telic and Agentive Roles from Corpus Data

Automatic Discovery of Telic and Agentive Roles from Corpus Data Ichiro Yamada Science & Technical Research Laboratories Japan Broadcasting Corporation 1-10-11 Kinuta, Setagaya-ku, Tokyo 157-8510, Japan yamada.i-hy@nhk.or.jp Timothy Baldwin CSSE University of Melbourne Victoria 3010, Australia tim@csse.unimelb.edu.au Abstract We present two methods for automatically discovering the telic and agentive roles of nouns from corpus data. These relations form part of the qualia structure assumed in generative lexi- con theory, where the telic role represents a typical purpose of the entity and the agentive role represents the origin of the entity. The first discovery method uses hand-generated templates for each role type, and the second employs a supervised machine-learning technique. To eval- uate the effectiveness of the two methods, we took a sample of 30 nouns, selected 50 verbs for each, and then generated a ranked list of verbs for a given noun. Using a variant of Spearman’s rank correlation, we demonstrate the ability of the methods to identify qualia structure. 1 Introduction We present a study of methods for automatically discovering the telic and agentive roles of nouns based on corpus data. These relations form part of the qualia structure assumed in generative lexicon theory (Pustejovsky, 1995). The qualia structure of a given noun incorporates (at most) the following four roles: � Formal role: the conceptual superclass of the noun. e.g., orientation, magnitude, shape, dimensionality, color, or position. � Constitutive role: the internal constitution of the entity. e.g., material, weight, parts, or component elements. � Telic role: the typical function of the entity. i.e. what the entity is for. � Agentive role: the origin of the entity, or its coming into being e.g., creator, artifact, natural kind, or casual chain For example, for the noun book, publication is a formal role noun, text is a constitutive role noun, read is a telic role verb, and write is an agentive role verb. Research has been done on extracting the formal and constitutive roles of nouns. Hearst (1992), Widdows and Dorow (2002), and others developed methods of automatically acquiring noun hyponyms— corresponding to the formal role—by identifying a set of frequently used and unambiguous lexico- syntactic patterns. Girju et al. (2003) proposed a method of learning part–whole relations, which cor- respond to the constitutive role. It is also possible to use lexical resources such as WordNet (Fellbaum, 1998) to determine formal (through hypernym links) and constitutive role data (through meronym links). Telic and agentive roles, on the other hand, have received relatively little attention in terms of automatic acquisition and are not available from any large-scale lexical resources. The only work we are aware of which directly targets the task of learning telic and agentive qualia data is that of Bouillon et al. (2002), PACLIC 18, December 8th-10th, 2004, Waseda University, Tokyo - 115 - who use symbolic learning to identify “qualia pairs”—token instances of noun–verb pairs which corre- spond to a some qualia role—in corpus data. Our work differs in that we can identify the qualia roles of an arbitrary noun, …

Content truncated to 500 words