Verb Phrase Ellipsis Resolution Using Discriminative and Margin-Infused Algorithms

Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 1734–1743, Austin, Texas, November 1-5, 2016. c⃝2016 Association for Computational Linguistics Verb Phrase Ellipsis Resolution Using Discriminative and Margin-Infused Algorithms Kian Kenyon-Dean Jackie Chi Kit Cheung Doina Precup School of Computer Science McGill University kian.kenyon-dean@mail.mcgill.ca, {jcheung,dprecup}@cs.mcgill.ca Abstract Verb Phrase Ellipsis (VPE) is an anaphoric construction in which a verb phrase has been elided. It occurs frequently in dialogue and informal conversational settings, but despite its evident impact on event coreference reso- lution and extraction, there has been relatively little work on computational methods for iden- tifying and resolving VPE. Here, we present a novel approach to detecting and resolving VPE by using supervised discriminative ma- chine learning techniques trained on features extracted from an automatically parsed, pub- licly available dataset. Our approach yields state-of-the-art results for VPE detection by improving F1 score by over 11%; additionally, we explore an approach to antecedent identifi- cation that uses the Margin-Infused-Relaxed- Algorithm, which shows promising results. 1 Introduction Verb Phrase Ellipsis (VPE) is an anaphoric construc- tion in which a verbal constituent has been omitted. In English, an instance of VPE consists of two parts: a trigger, typically an auxiliary or modal verb, that indicates the presence of a VPE; and an antecedent, which is the verb phrase to which the elided element resolves (Bos and Spenader, 2011; Dalrymple et al., 1991). For example, in the sentence, “The govern- ment includes money spent on residential renova- tion; Dodge does not”, the trigger “does” resolves to the antecedent “includes money spent on residen- tial renovation”. The ability to perform VPE resolution is impor- tant for tasks involving event extraction, especially in conversational genres such as informal dialogue where VPE occurs more frequently (Nielsen, 2005). Most current event extraction systems ignore VPE and derive some structured semantic representation by reading information from a shallow dependency parse of a sentence. Such an approach would not only miss many valid links between an elided verb and its arguments, it could also produce nonsensi- cal extractions if applied directly on an auxiliary trigger. In the example above, a naive approach might produce an unhelpful semantic triple such as (Dodge, agent, do). There have been several previous empirical stud- ies of VPE (Hardt, 1997; Nielsen, 2005; Bos and Spenader, 2011; Bos, 2012; Liu et al., 2016). Many previous approaches were restricted to solving spe- cific subclasses of VPE (e.g., VPE triggered by do (Bos, 2012)), or have relied on simple heuristics for some or all of the steps in VPE resolution, such as by picking the most recent previous clause as the an- tecedent. In this paper, we develop a VPE resolution pipeline which encompasses a broad class of VPEs (Figure 1), decomposed into the following two steps. In the VPE detection step, the goal is to determine whether or not a word triggers VPE. The second step, antecedent identification, requires selecting the clause containing the verbal antecedent, as well as determining the exact boundaries of …

Content truncated to 500 words