A Supervised Algorithm for Verb Disambiguation into VerbNet Classes

Proceedings of the 22nd International Conference on Computational Linguistics (Coling 2008), pages 9–16 Manchester, August 2008 A Supervised Algorithm for Verb Disambiguation into VerbNet Classes Omri Abend1 Roi Reichart2 Ari Rappoport1 1Institute of Computer Science , 2ICNC Hebrew University of Jerusalem {omria01|roiri|arir}@cs.huji.ac.il Abstract VerbNet (VN) is a major large-scale En- glish verb lexicon. Mapping verb instances to their VN classes has been proven use- ful for several NLP tasks. However, verbs are polysemous with respect to their VN classes. We introduce a novel supervised learning model for mapping verb instances to VN classes, using rich syntactic features and class membership constraints. We evaluate the algorithm in both in-domain and corpus adaptation scenarios. In both cases, we use the manually tagged Sem- link WSJ corpus as training data. For in- domain (testing on Semlink WSJ data), we achieve 95.9% accuracy, 35.1% error re- duction (ER) over a strong baseline. For adaptation, we test on the GENIA corpus and achieve 72.4% accuracy with 10.7% ER. This is the first large-scale experimen- tation with automatic algorithms for this task. 1 Introduction The organization of verbs into classes whose mem- bers exhibit similar syntactic and semantic behav- ior has been discussed extensively in the linguistics literature (see e.g. (Levin and Rappaport Hovav, 2005; Levin, 1993)). Such an organization helps in avoiding lexicon representation redundancy and enables generalizations across similar verbs. It can also be of great practical use, e.g. in com- pensating NLP statistical models for data sparse- ness. Indeed, Levin’s seminal work had motivated c⃝ 2008. Licensed under the Creative Commons Attribution-Noncommercial-Share Alike 3.0 Unported li- cense (http://creativecommons.org/licenses/by-nc-sa/3.0/). Some rights reserved. much research aimed at automatic discovery of verb classes (see Section 2). VerbNet (VN) (Kipper et al., 2000; Kipper- Schuler, 2005) is a large scale, publicly available domain independent verb lexicon that builds on Levin classes and extends them with new verbs, new classes, and additional information such as semantic roles and selectional restrictions. VN classes were proven beneficial for Semantic Role Labeling (SRL) (Swier and Stevenson, 2005), Se- mantic Parsing (Shi and Mihalcea, 2005) and building conceptual graphs (Hensman and Dun- nion, 2004). Levin-inspired classes have been used in several NLP tasks, such as Machine Trans- lation (Dorr, 1997) and Document Classification (Klavans and Kan, 1998). Many applications that use VN need to map verb instances onto their VN classes. However, verbs are polysemous with respect to VN classes. Sem- link (Loper et al., 2007) is a dataset that maps each verb instance in the WSJ Penn Treebank to its VN class. The mapping has been created using a com- bination of automatic and manual methods. Yi et al. (2007) have used Semlink to improve SRL. In this paper we present the first large-scale ex- perimentation with a supervised machine learning classification algorithm for disambiguating verb instances to their VN classes. We use rich syntactic features extracted from a treebank-style parse tree, and utilize a learning algorithm capable of impos- ing class membership constraints, thus taking ad- vantage of the nature …

Content truncated to 500 words