A Strategic Reasoning Model for Generating Alternative Answers

Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing, pages 534–542, Beijing, China, July 26-31, 2015. c⃝2015 Association for Computational Linguistics A Strategic Reasoning Model for Generating Alternative Answers Jon Scott Stevens Center for General Linguistics, Berlin stevens@zas.gwz-berlin.de Anton Benz Center for General Linguistics, Berlin Sebastian Reuße Ruhr University Bochum sebastian.reusse@rub.de Ralf Klabunde Ruhr University Bochum Abstract We characterize a class of indirect an- swers to yes/no questions, alternative an- swers, where information is given that is not directly asked about, but which might nonetheless address the underlying moti- vation for the question. We develop a model rooted in game theory that gener- ates these answers via strategic reasoning about possible unobserved domain-level user requirements. We implement the model within an interactive question an- swering system simulating real estate dia- logue. The system learns a prior probabil- ity distribution over possible user require- ments by analyzing training dialogues, which it uses to make strategic deci- sions about answer selection. The system generates pragmatically natural and inter- pretable answers which make for more ef- ficient interactions compared to a baseline. 1 Introduction In natural language dialogue, questions are often answered indirectly. This is particularly apparent for yes/no questions, where a wide range of re- sponses beyond literal “yes” and “no” answers is available. Sometimes indirect answers serve to an- ticipate the next step of the hearer’s plan, as in (1) (Allen and Perrault, 1980), where the literal an- swer is entailed by the supplied answer, and some- times indirect answers leave it to the hearer to in- fer the literal answer from common contextual as- sumptions, as in (2) (de Marneffe et al., 2009). (1) Q: Has the train to Windsor left yet? A: It’s leaving soon from gate 7. (2) Q: Is Sue at work? A: She’s sick with the flu. But other times there is no semantic link between the question and the supplied answer. Rather, the answer must be interpreted in light of the task- specific goals of the interlocutors. Consider (3) in a context where a customer is posing questions to a real estate agent with the aim of renting an apartment. (3) Q: Does the apartment have a garden? A: Well, it has a large balcony. Whether there is a balcony has no logical bear- ing on whether there is a garden. Intuitively, the realtor is inferring that the customer’s question might have been motivated by a more general re- quirement (perhaps the customer wants a place to grow flowers) and supplying an alternative at- tribute to satisfy that requirement. In this case the answerer must reason about which attributes of an apartment might satisfy a customer who would ask about a garden. Note that multiple motivating requirements are possible (perhaps the customer just wants to relax outside), such that the answerer might just as easily have said, “It has a large bal- cony, and there is a park close by.” In either …

Content truncated to 500 words