Files in this item



application/pdfJohn_Wieting.pdf (260kB)
(no description provided)PDF


Title:Lexical entailment
Author(s):Wieting, John
Advisor(s):Roth, Dan
Department / Program:Computer Science
Discipline:Computer Science
Degree Granting Institution:University of Illinois at Urbana-Champaign
Subject(s):Lexical Entailment
Recognizing Textual Entailment
Word Similarity
Word Representations
Natural Language Processing
Abstract:Lexical entailment is a requirement for success in the domains of Recognizing Textual Entailment (RTE) as well as related tasks like Question-Answering and Information Extraction. Previous approaches tend to fall into two camps - those that make use of distributional models and those that make use of knowledge bases such as WordNet. Interestingly, these methods make very different kinds of mistakes and so in this thesis, we construct a new entailment measure by combining these two paradigms in such a way that exploits their differences. We also experiment with including local context and modify an existing approach to achieve the best unsupervised performance so far on the Lexical Substitution task. Overall, we achieve a significant gain in performance on three different evaluations and our approach is also faster than the other distributional approaches we compare to as we are able to avoid fruitless comparisons. Furthermore, we introduce a new approach to evaluate lexical entailment that avoids some of the issues of wordlists - the current conventional way of evaluating lexical entailment, and that can be additionally used to evaluate lexical entailment in context - a novel task introduced in this paper. We also include experiments that show our new lexical entailment model improves performance on the RTE task, the main goal of this work.
Issue Date:2014-05-30
Rights Information:Copyright 2014 John Wieting
Date Available in IDEALS:2014-05-30
Date Deposited:2014-05

This item appears in the following Collection(s)

Item Statistics