Resolving elliptical questions in Question Answering Systems
Nuria Bertomeu

For NLQA Systems to be really user friendly, they must be able to successfully resolve elliptical questions, since users do not normally ask isolated questions, but make small researches about a particular topic, thus formulating several questions, usually related to each other. These questions, though being incomplete at their outward form, have a full meaning in context.
In the literature there have been several approaches to ellipsis resolution, but they have mostly concentrated on the resolution of VP-ellipsis, i.e., HOU (Dalrymple, Shieber and Pereira, 1991), substitution of quasi logical forms (Alshawi et al., 1992), among others; and on the resolution of fragmentary answers and clarification utterances (Ginzburg 1996, 2001). Carbonell (1983), on the other hand, proposes a case-based strategy to the resolution of elliptical questions, but its use is limited to natural language interfaces to databases and requires rather domain-specific semantic information. In this talk I’m going to take a look at some kinds of elliptical questions, in order to see how they differ from other cases of ellipsis, especially regarding the focus-ground articulation, and at whether existing theoretical approaches to ellipsis resolution can account for them and how.
I’m also going to present two approaches to the practical resolution of elliptical questions, which I expect to develop during my research and whose viability for the use in practical QA systems is set here as an empirical question. One of them involves using  the DeepThought core architecture framework "Heart of Gold" (Callmeier et al., 2003), which provides an infrastructure for combining different types of NLP modules, deep and shallow, and their information output through a uniform representation language, XML-encoded RMRS, to analyze the  alone-standing fragments and integrate the output in the more accurate and deep representation of the sentence being the source of the ellipsis. The other one is to develop a piece of grammar to account for elliptical questions which would allow the deep parser to deliver analysis of elliptical fragments. Here I will depart from the strategy developed within the framework for dialogue representation KOS (Ginzburg 1996, 2001; Bohlin et al. 1999; Larsson 2002), which proposes, within the framework of HPSG (Pollard & Sag, 1994; Sag & Wasow, 1999), an account of elliptical fragments, concretely answers to wh-questions and clarification utterances, using contextual information, and which has been successfully implemented in SHARDS (Ginzburg & Fernández, 2004).