Unsupervised Discovery of a Statistical Verb Lexicon

Proceedings of the 2006 Conference on Empirical Methods in Natural Language Processing (EMNLP 2006), pages 1–8, Sydney, July 2006. c⃝2006 Association for Computational Linguistics Unsupervised Discovery of a Statistical Verb Lexicon Trond Grenager and Christopher D. Manning Computer Science Department Stanford University Stanford, CA 94305 {grenager, manning}@cs.stanford.edu Abstract This paper demonstrates how unsupervised tech- niques can be used to learn models of deep linguis- tic structure. Determining the semantic roles of a verb’s dependents is an important step in natural language understanding. We present a method for learning models of verb argument patterns directly from unannotated text. The learned models are sim- ilar to existing verb lexicons such as VerbNet and PropBank, but additionally include statistics about the linkings used by each verb. The method is based on a structured probabilistic model of the do- main, and unsupervised learning is performed with the EM algorithm. The learned models can also be used discriminatively as semantic role labelers, and when evaluated relative to the PropBank anno- tation, the best learned model reduces 28% of the error between an informed baseline and an oracle upper bound. 1 Introduction An important source of ambiguity that must be resolved by any natural language understanding system is the mapping between syntactic depen- dents of a predicate and the semantic roles1 that they each express. The ambiguity stems from the fact that each predicate can allow several alternate mappings, or linkings,2 between its semantic roles and their syntactic realization. For example, the verb increase can be used in two ways: (1) The Fed increased interest rates. (2) Interest rates increased yesterday. The instances have apparently similar surface syn- tax: they both have a subject and a noun phrase directly following the verb. However, while the subject of increase expresses the agent role in the first, it instead expresses the patient role in the sec- ond. Pairs of linkings such as this allowed by a single predicate are often called diathesis alterna- tions (Levin, 1993). The current state-of-the-art approach to resolv- ing this ambiguity is to use discriminative classi- fiers, trained on hand-tagged data, to classify the 1Also called thematic roles, theta roles, or deep cases. 2Sometimes called frames. semantic role of each dependent (Gildea and Juraf- sky, 2002; Pradhan et al., 2005; Punyakanok et al., 2005). A drawback of this approach is that even a relatively large training corpus exhibits consid- erable sparsity of evidence. The two main hand- tagged corpora are PropBank (Palmer et al., 2003) and FrameNet (Baker et al., 1998), the former of which currently has broader coverage. However, even PropBank, which is based on the 1M word WSJ section of the Penn Treebank, is insufficient in quantity and genre to exhibit many things. A perfectly common verb like flap occurs only twice, across all morphological forms. The first example is an adjectival use (flapping wings), and the sec- ond is a rare intransitive use with an agent argu- ment and a path (ducks flapping over Washington). From this data, one cannot learn …

Content truncated to 500 words