A General Feature Space for Automatic Verb Classification

A General Feature Space for Automatic Verb Classification Eric Joanis and Suzanne Stevenson Department of Computer Science University of Toronto fjoanis,suzannel@cs.toronto.edu Abstract We develop a general feature space for automatic classification of verbs into lexical semantic classes. Previous work was limited in scope by the need for manual selection of discriminating fea- tures, through a linguistic analysis of the target verb classes (Merlo and Steven- son, 2001). We instead analyze the classification structure at a higher level, using the possible defining characteris- tics of classes as the basis for our fea- ture space. The general feature space achieves reductions in error rates of 42— 69%, on a wider range of classes than investigated previously, with compara- ble performance to feature sets manu- ally selected for the particular classifi- cation tasks. Our results show that the approach is generally applicable, and avoids the need for resource-intensive linguistic analysis for each new task. 1 Introduction Wide-coverage language processing systems re- quire large amounts of knowledge about individ- ual words, leading to a lexical acquisition bottle- neck. Because verbs play a central role in the syn- tactic and semantic interpretation of a sentence, much research has focused on automatically learn- ing properties of verbs from text corpora, such as their subcategorization (Brent, 1993; Briscoe and Carroll, 1997), argument roles (Riloff and Schmelzenbach, 1998; Gildea and Jurafsky, 2002), selectional preferences (Resnik, 1996), and lexi- cal semantic classification (Dorr and Jones, 1996; Lapata and Brew, 1999; Schulte im Walde, 2000; Merlo and Stevenson, 2001). Our work aims to ex- tend the applicability of the latter, by developing a general feature space for automatic verb classifi- cation. Specifically, Merlo and Stevenson (2001) showed that verbs could be automatically classi- fied into one of three lexical semantic classes on the basis of five simple statistical features. This work demonstrated the feasibility of verb classifi- cation from noisy, easily extractable corpus statis- tics. However, the approach was limited in scope by the need for manual determination of a discrim- inating set of features, through a linguistic anal- ysis of the target verb classes. Our work over- comes this limitation by developing a general fea- ture space that avoids the need for individual de- velopment of features for specific classes. The central idea of our approach is as follows. We focus on lexical semantic classes as in Levin (1993) (hereafter Levin), which group together verbs sharing both a common semantics (such as manner of motion or change of state), and a set of syntactic alternations. An alternation refers to the alternative mappings of the semantic arguments of a verb to syntactic positions, as in: la. I loaded the truck b. I loaded hay onto the truck. In Merlo and Stevenson (2001) (hereafter MS01), the differences between the specific semantic ar- guments and their possible alternations for the three target classes were analyzed to determine a small set of discriminating features. Here, we per- form the linguistic analysis at a higher level, by analyzing the range of possible alternations and …

Content truncated to 500 words