Leveraging VerbNet to build Corpus-Specific Verb Clusters

Proceedings of the Fifth Joint Conference on Lexical and Computational Semantics (*SEM 2016), pages 102–107, Berlin, Germany, August 11-12, 2016. Leveraging VerbNet to build Corpus-Specific Verb Clusters Daniel W Peterson and Jordan Boyd-Graber and Martha Palmer University of Colorado {daniel.w.peterson,jordan.boyd.graber,martha.palmer}@colorado.edu Daisuke Kawhara Kyoto University, JP dk@i.kyoto-u.ac.jp Abstract In this paper, we aim to close the gap from extensive, human-built semantic re- sources and corpus-driven unsupervised models. The particular resource explored here is VerbNet, whose organizing princi- ple is that semantics and syntax are linked. To capture patterns of usage that can aug- ment knowledge resources like VerbNet, we expand a Dirichlet process mixture model to predict a VerbNet class for each sense of each verb, allowing us to incorpo- rate annotated VerbNet data to guide the clustering process. The resulting clusters align more closely to hand-curated syn- tactic/semantic groupings than any previ- ous models, and can be adapted to new domains since they require only corpus counts. 1 Introduction In this paper, we aim to close the gap from exten- sive, human-built semantic resources and corpus- driven unsupervised models. The work done by linguists over years of effort has been validated by the scientific community, and promises real trac- tion on the fuzzy problem of deriving meaning from words. However, lack of coverage and adapt- ability currently limit the usefulness of this work. The particular resource explored here is Verb- Net (Kipper-Schuler, 2005), a semantic resource built upon the foundation of verb classes by Levin (1993). Levin’s verb classes are built on the hy- pothesis that syntax and semantics are fundamen- tally linked. The semantics of a verb affect the allowable syntactic constructions involving that verb, creating regularities in language to which speakers are extremely sensitive. It follows that grouping verbs by allowable syntactic realizations leads from syntax to meaningful semantic group- ings. This seed grew into VerbNet, a process which involved dozens of linguists and a decade of work, making careful decisions about the al- lowable syntactic frames for various verb senses, informed by text examples. VerbNet is useful for semantic role labeling and related tasks (Giuglea and Moschitti, 2006; Yi, 2007; Yi et al., 2007; Merlo and van der Plas, 2009; Kshirsagar et al., 2014), but its widespread use is limited by coverage. Not all verbs have a VerbNet class, and some polysemous verbs have important senses unaccounted for. In addi- tion, VerbNet is not easily adaptable to domain- specific corpora, so these omissions may be more prominent outside of the general-purpose corpora and linguistic intuition used in its construction. Its great strength is also its downfall: adding new verbs, new senses, and new classes requires trained linguists - at least, to preserve the integrity of the resource. According to Levin’s hypothesis, knowing the set of allowable syntactic patterns for a verb sense is sufficient to make meaningful semantic classifi- cations. Large-scale corpora provide an extremely comprehensive picture of the possible syntactic re- alizations for any particular verb. With enough data in the training set, even infrequent verbs …

Content truncated to 500 words