Improved Lexical Acquisition through DPP-based Verb Clustering

Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics, pages 862–872, Sofia, Bulgaria, August 4-9 2013. c⃝2013 Association for Computational Linguistics Improved Lexical Acquisition through DPP-based Verb Clustering Roi Reichart University of Cambridge, UK rr439@cam.ac.uk Anna Korhonen University of Cambridge, UK alk23@cam.ac.uk Abstract Subcategorization frames (SCFs), selec- tional preferences (SPs) and verb classes capture related aspects of the predicate- argument structure. We present the first unified framework for unsupervised learn- ing of these three types of information. We show how to utilize Determinantal Point Processes (DPPs), elegant proba- bilistic models that are defined over the possible subsets of a given dataset and give higher probability mass to high qual- ity and diverse subsets, for clustering. Our novel clustering algorithm constructs a joint SCF-DPP DPP kernel matrix and uti- lizes the efficient sampling algorithms of DPPs to cluster together verbs with sim- ilar SCFs and SPs. We evaluate the in- duced clusters in the context of the three tasks and show results that are superior to strong baselines for each 1. 1 Introduction Verb classes (VCs), subcategorization frames (SCFs) and selectional preferences (SPs) capture different aspects of predicate-argument structure. SCFs describe the syntactic realization of verbal predicate-argument structure, SPs capture the se- mantic preferences verbs have for their arguments and VCs in the Levin (1993) tradition provide a shared level of abstraction for verbs that share many aspects of their syntactic and semantic be- havior. These three of types of information have proved useful for Natural Language Processing (NLP) 1The source code of the clustering algorithms and evalu- ation is submitted with this paper and will be made publicly available upon acceptance of the paper. tasks which require information about predicate- argument structure, including parsing (Shi and Mi- halcea, 2005; Cholakov and van Noord, 2010; Zhou et al., 2011), semantic role labeling (Swier and Stevenson, 2004; Dang, 2004; Bharati et al., 2005; Moschitti and Basili, 2005; zap, 2008; Zapi- rain et al., 2009), and word sense disambiguation (Dang, 2004; Thater et al., 2010; ´O S´eaghdha and Korhonen, 2011), among many others. Because lexical information is highly sensitive to domain variation, approaches that can identify VCs, SCFs and SPs in corpora have become in- creasingly popular, e.g. (O’Donovan et al., 2005; Schulte im Walde, 2006; Erk, 2007; Preiss et al., 2007; Van de Cruys, 2009; Reisinger and Mooney, 2011; Sun and Korhonen, 2011; Lippincott et al., 2012). The task of SCF induction involves identifying the arguments of a verb lemma and generalizing about the frames (i.e. SCFs) taken by the verb, where each frame includes a number of arguments and their syntactic types. For example, in (1), the verb ”show” takes the frame SUBJ-DOBJ- CCOMP (subject, direct object, and clausal complement). (1) [A number of SCF acquisition papers]SUBJ [show]VERB [their readers]DOBJ [which fea- tures are most valuable for the acquisition process]CCOMP. SP induction involves identifying and classify- ing the lexical items in a given argument slot. In sentence (2), for example, the verb ”show” takes the frame SUBJ-DOBJ. The …

Content truncated to 500 words