Antecedent Selection for Sluicing: Structure and Content

Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 1234–1243, Austin, Texas, November 1-5, 2016. c⃝2016 Association for Computational Linguistics Antecedent Selection for Sluicing: Structure and Content Pranav Anand Linguistics UC Santa Cruz panand@ucsc.edu Daniel Hardt IT Management Copenhagen Business School dh.itm@cbs.dk Abstract Sluicing is an elliptical process where the ma- jority of a question can go unpronounced as long as there is a salient antecedent in previ- ous discourse. This paper considers the task of antecedent selection: finding the correct antecedent for a given case of sluicing. We argue that both syntactic and discourse rela- tionships are important in antecedent selec- tion, and we construct linguistically sophis- ticated features that describe the relevant re- lationships. We also define features that de- scribe the relation of the content of the an- tecedent and the sluice type. We develop a lin- ear model which achieves accuracy of 72.4%, a substantial improvement over a strong man- ually constructed baseline. Feature analysis confirms that both syntactic and discourse fea- tures are important in antecedent selection. 1 Introduction Ellipsis involves sentences with missing subparts, where those subparts must be interpretatively filled in by the hearer. How this is possible has been a ma- jor topic in linguistic theory for decades (Sag, 1976; Chung et al., 1995; Merchant, 2001). One widely studied example is verb phrase ellipsis (VPE), ex- emplified by (1). (1) Harry traveled to southern Denmark to study botany . Tom did too . In the second sentence (Tom did too) the verb phrase is entirely missing, yet the hearer effortlessly ‘re- solves’ (understands) its content to be traveled to southern Denmark to study botany. Another widely studied case of ellipsis is sluicing, in which the majority of a question is unpronounced, as in (2). (2) Harry traveled to southern Denmark to study botany . I want to know why . Here the content of the question, introduced by the WH-phrase why, is missing, yet it is understood by the hearer to be why did Harry travel to southern Denmark to study botany?. In both of these cases, ellipsis resolution is made possible by the presence of an antecedent, material in prior discourse that, in- formally speaking, is equivalent to what is missing. Ellipsis poses an important challenge for many applications in language technology, as various forms of ellipsis are known to be frequent in a va- riety of languages and text types. This is perhaps most evident in the case of question-answering sys- tems, since elliptical questions and elliptical answers are both very common in discourse. A computa- tional system that can effectively deal with ellipsis involves three subtasks (Nielsen, 2005): ellipsis de- tection, in which a case of ellipsis is identified, an- tecedent selection, in which the antecedent for a case of ellipsis is found, and ellipsis resolution, where the content of the ellipsis is filled in with reference to the antecedent and the context of the ellipsis. Here, we focus on antecedent selection for sluicing. In …

Content truncated to 500 words