Hierarchical Verb Clustering Using Graph Factorization

Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing, pages 1023–1033, Edinburgh, Scotland, UK, July 27–31, 2011. c⃝2011 Association for Computational Linguistics Hierarchical Verb Clustering Using Graph Factorization Lin Sun and Anna Korhonen University of Cambridge, Computer Laboratory 15 JJ Thomson Avenue, Cambridge CB3 0GD, UK ls418,alk23@cl.cam.ac.uk Abstract Most previous research on verb clustering has focussed on acquiring flat classifications from corpus data, although many manually built classifications are taxonomic in nature. Also Natural Language Processing (NLP) applica- tions benefit from taxonomic classifications because they vary in terms of the granularity they require from a classification. We intro- duce a new clustering method called Hierar- chical Graph Factorization Clustering (HGFC) and extend it so that it is optimal for the task. Our results show that HGFC outperforms the frequently used agglomerative clustering on a hierarchical test set extracted from VerbNet, and that it yields state-of-the-art performance also on a flat test set. We demonstrate how the method can be used to acquire novel classifi- cations as well as to extend existing ones on the basis of some prior knowledge about the classification. 1 Introduction A variety of verb classifications have been built to support NLP tasks. These include syntactic and se- mantic classifications, as well as ones which in- tegrate aspects of both (Grishman et al., 1994; Miller, 1995; Baker et al., 1998; Palmer et al., 2005; Kipper, 2005; Hovy et al., 2006). Classifications which integrate a wide range of linguistic proper- ties can be particularly useful for tasks suffering from data sparseness. One such classification is the taxonomy of English verbs proposed by Levin (1993) which is based on shared (morpho-)syntactic and semantic properties of verbs. Levin’s taxon- omy or its extended version in VerbNet (Kipper, 2005) has proved helpful for various NLP applica- tion tasks, including e.g. parsing, word sense disam- biguation, semantic role labeling, information ex- traction, question-answering, and machine transla- tion (Swier and Stevenson, 2004; Dang, 2004; Shi and Mihalcea, 2005; Zapirain et al., 2008). Because verbs change their meaning and be- haviour across domains, it is important to be able to tune existing classifications as well to build novel ones in a cost-effective manner, when required. In recent years, a variety of approaches have been pro- posed for automatic induction of Levin style classes from corpus data which could be used for this pur- pose (Schulte im Walde, 2006; Joanis et al., 2008; Sun et al., 2008; Li and Brew, 2008; Korhonen et al., 2008; ´O S´eaghdha and Copestake, 2008; Vla- chos et al., 2009). The best of such approaches have yielded promising results. However, they have mostly focussed on acquiring and evaluating flat classifications. Levin’s classification is not flat, but taxonomic in nature, which is practical for NLP pur- poses since applications differ in terms of the gran- ularity they require from a classification. In this paper, we experiment with hierarchical Levin-style clustering. We adopt as our baseline method a well-known hierarchical method – ag- glomerative clustering (AGG) – which has …

Content truncated to 500 words