Date of Award

Spring 1-1-2019

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

First Advisor

Martha Palmer

Second Advisor

James Martin

Third Advisor

Daisuke Kawahara

Fourth Advisor

Claire Monteleoni

Fifth Advisor

Mans Hulden

Abstract

The objective of this research is to build automated models that emulate VerbNet, a semantic resource for English verbs. VerbNet has been built and expanded by linguists, forming a hierarchical clustering of verbs with common semantic and syntactic expressions, and is useful in semantic tasks. A major drawback is the difficulty of extending a manually-curated resource, which leads to gaps in coverage. After over a decade of development, VerbNet has missing verbs, missing senses of common verbs, and is missing appropriate classes to contain at least some of them. Although there have been efforts to build VerbNet resources in other languages, none have received as much attention, so these coverage issues are often more glaring in resource-poor languages. Probabilistic models can emulate VerbNet by learning distributions from large corpora, addressing coverage by providing both a complete clustering of the observed data, and a model to assign unseen sentences to clusters. The output of these models can aid the creation and expansion of VerbNet in English and other languages, especially if they align strongly with known VerbNet classes.

This work develops several improvements to the state-of-the-art system for verb sense induction and VerbNet-like clustering. The baseline is two-step process for automatically inducing verb senses and producing a polysemy-aware clustering, that matched VerbNet more closely than any previous methods. First, we will see that a single-step process can produce better automatic senses and clusters. Second, we explore an alternative probabilistic model, which is successful on the verb clustering task. This model does not perform well on sense induction, so we analyze the limitations on its applicability. Third, we explore methods of supervising these probabilistic models with limited labeled data, which dramatically improves the recovery of correct clusters. Together these improvements suggest a line of research for practitioners to take advantage of probabilistic models in VerbNet annotation efforts.

Share

COinS