Unsupervised Induction of a Syntax-Semantics Lexicon Using Iterative Refinement

First Joint Conference on Lexical and Computational Semantics (*SEM), pages 180–188, Montr´eal, Canada, June 7-8, 2012. c⃝2012 Association for Computational Linguistics Unsupervised Induction of a Syntax-Semantics Lexicon Using Iterative Refinement Hagen F¨urstenau CCLS Columbia University New York, NY, USA hagen@ccls.columbia.edu Owen Rambow CCLS Columbia University New York, NY, USA rambow@ccls.columbia.edu Abstract We present a method for learning syntax- semantics mappings for verbs from unanno- tated corpora. We learn linkings, i.e., map- pings from the syntactic arguments and ad- juncts of a verb to its semantic roles. By learn- ing such linkings, we do not need to model in- dividual semantic roles independently of one another, and we can exploit the relation be- tween different mappings for the same verb, or between mappings for different verbs. We present an evaluation on a standard test set for semantic role labeling. 1 Introduction A verb can have several ways of mapping its seman- tic arguments to syntax (“diathesis alternations”): (1) a. We increased the response rate with SHK. b. SHK increased the response rate. c. The response rate increased. The subject of increase can be the agent (1a), the in- strument (1b), or the theme (what is being increased) (1c). Other verbs that show this pattern include break or melt. Much theoretical and lexicographic (descriptive) work has been devoted to determining how verbs map their lexical predicate-argument structure to syntactic arguments (Burzio, 1986; Levin, 1993). The last decades have seen a surge in activity on the computational front, spurred in part by efforts to annotate large corpora for lexical semantics (Baker et al., 1998; Palmer et al., 2005). Initially, we have seen computational efforts devoted to finding classes of verbs that share similar syntax-semantics map- pings from annotated and unannotated corpora (La- pata and Brew, 1999; Merlo and Stevenson, 2001). More recently, there has been an explosion of inter- est in semantic role labeling (with too many recent publications to cite). In this paper, we explore learning syntax- semantics mappings for verbs from unannotated cor- pora. We are specifically interested in learning link- ings. A linking is a mapping for one verb from its syntactic arguments and adjuncts to all of its se- mantic roles, so that individual semantic roles are not modeled independently of one another and so that we can exploit the relation between different mappings for the same verb (as in (1) above), or between mappings for different verbs. We there- fore follow Grenager and Manning (2006) in treat- ing linkings as first-class objects; however, we dif- fer from their work in two important respects. First, we use semantic clustering of head words of argu- ments in an approach that resembles topic modeling, rather than directly modeling the subcategorization of verbs with a distribution over words. Second and most importantly, we do not make any assumptions about the linkings, as do Grenager and Manning (2006). They list a small set of rules from which they derive all linkings possible in their model; in contrast, we are able to learn any …

Content truncated to 500 words