Despite a plethora of recent work on questions across linguistics, logic, and computer science, there is a communication gap between work on questions in different fields. In research on discourse there is attention to the wide variety of responses that can be made to questions, and the role of implicit questions in structuring discourse (e.g. Roberts 1996, Ginzburg 2010, Farkas and Bruce 2010. In semantics, the focus is often on embedded questions, and therefore on highly restricted notions of (Karttunen 1977 and much subsequent literature).
In parallel to these, an active computational literature on question-answering (QA) as a practical task has emerged that is highly successful in constrained practical domains, but rarely makes reference to any of the extant semantic or pragmatic theories (e.g. Ferrucci et al 2010, Yao and Van Durme 2014). QA systems such as IBM’s Watson tend to operate in highly constrained discourse settings: Jeopardy discourses are asymmetric, 2-party, non-interactive, and involve relatively short, typically factual questions with a single correct answer. This kind of problem is already incredibly complex, and systems such as Watson (or research systems such as Yao and Van Durme’s 2014 Freebase-based system) now do quite well within such constraints. However, these tasks are far from naturalistic.
The goal of this class is to explore how tools from formal pragmatics might connect to the computational setting, and what challenges naturalistic discourse introduces to automatic QA, as well as to explore what linguistics researchers might learn from successful QA systems.