Modeling human clarification question production through expected regret

Polina Tsvilodub, Michael Franke, Robert D. Hawkins

Keywords: clarification question asking, computational pragmatics, expected regret, goal uncertainty

TL;DR: We provide empirical results and an expected regret based computational model for human choices to produce clarification questions under uncertainty.

Abstract: Wh-questions are ubiquitous in everyday conversations, and there are many ways in which speakers respond to them. Decision-theoretic accounts of question answering in pragmatics posit that cooperative answerers respond so as to maximize the questioner’s expected utility with respect to their latent goal, but when faced with uncertainty about that goal, the answerer might choose to ask a (clarification) question. We propose that answerers navigate this decision through expected regret: they consider how much they might regret giving any available answer (i.e., exhaustive or mention-some) if wrong about the questioner’s goal, and ask for clarification when expected regret is high. We find that that participants in an online experiment prefer asking clarification questions under high uncertainty about the questioner’s goal when the exhaustive answer is costly, and that our model can capture this empirical response pattern.

Submission Number: 37