Exploiting a Verb Lexicon in Automatic Semantic Role Labelling

Proceedings of Human Language Technology Conference and Conference on Empirical Methods in Natural Language Processing (HLT/EMNLP), pages 883–890, Vancouver, October 2005. c⃝2005 Association for Computational Linguistics Exploiting a Verb Lexicon in Automatic Semantic Role Labelling Robert S. Swier and Suzanne Stevenson Department of Computer Science University of Toronto Toronto, Ontario, Canada M5S 3G4 {swier,suzanne}@cs.toronto.edu Abstract We develop an unsupervised semantic role labelling system that relies on the direct application of information in a predicate lexicon combined with a simple probabil- ity model. We demonstrate the usefulness of predicate lexicons for role labelling, as well as the feasibility of modifying an existing role-labelled corpus for evaluat- ing a different set of semantic roles. We achieve a substantial improvement over an informed baseline. 1 Introduction Intelligent language technologies capable of full semantic interpretation of domain-general text re- main an elusive goal. However, statistical advances have made it possible to address core pieces of the problem. Recent years have seen a wealth of research on one important component of seman- tic interpretation—automatic role labelling (e.g., Gildea and Jurafsky, 2002; Pradhan et al., 2004; Ha- cioglu et al., 2004, and additional papers from Car- reras and Marquez, 2004). Such work aims to an- notate each constituent in a clause with a semantic tag indicating the role that the constituent plays with respect to the target predicate, as in (1): (1) [Yuka]Agent [whispered]Pred to [Dar]Recipient Semantic role labelling systems address a crucial first step in the automatic extraction of semantic re- lations from domain-general text, taking us closer to the goal of comprehensive semantic mark-up. Most work thus far on domain-general role la- belling depends on supervised learning over statis- tical features extracted from a hand-labelled corpus. The reliance on such a resource—one in which the arguments of each predicate are manually identified and assigned a semantic role—limits the portability of such methods to other languages or even to other genres of corpora. In this study, we explore the possibility of using a verb lexicon, rather than a hand-labelled corpus, as the primary resource in the semantic role labelling task. Perhaps because of the focus on what can be gleaned from labelled data, existing supervised approaches have made little use of the additional knowledge available in the predicate lexicon asso- ciated with the labelled corpus. By contrast, we ex- ploit the explicit knowledge of the role assignment possibilities for each verb within an existing lexi- con. Moreover, we utilize a very simple probability model within a highly efficient algorithm. We use VerbNet (Kipper et al., 2000), a computa- tional lexicon which lists the possible semantic role assignments for each of its verbs. Our algorithm extracts automatically parsed arguments from a cor- pus, and assigns to each a list of the compatible roles according to VerbNet. Arguments which are given only a single role possibility are considered to have been assigned an unambiguous role label. This set of arguments constitutes our primary-labelled data, which serves as the noisy training data for a simple probability model which …

Content truncated to 500 words