The Alignment Model of Indirect Communication

Speakers often choose utterances under uncertainty about the potential opinions of the listener. In this case, utterances that do not signal the speaker's opinion directly may allow the speaker to avoid possible conflict. Adopting the Rational Speech Act framework, we develop a model of indirect communication that is able to (1) rationalize the choice of indirect utterances when speakers' opinions do not align; (2) capture complex reasoning about the interlocutor's opinion based on their reaction to an initial statement. The model has several novel features: the social utility component in the speaker function is computed by simulating potential divergences of opinions between conversation partners. This simulation becomes possible due to the inclusion of higher-order beliefs in the model machinery. The model further considers multi-turn dialogues rather than isolated utterances when inferring listener's opinion. We are therefore able to take a longer planning horizon into account when calculating the probability of utterances and use preceding discourse in social reasoning. The model, though complex, makes novel, non-trivial qualitative predictions, which are supported by data from three behavioral experiments reported here.