The alignment model of indirect communication

Abstract

Speakers often choose utterances under uncertainty about the potential opinion of the listener. In this case, utterances that do not signal the speaker’s opinion directly may allow the speaker to avoid possible conflict: saying that an election outcome is interesting rather than amazing, even if the speaker is truly excited about it, may give her an option to retreat if it turns out that the listener’s opinion is the opposite. By enhancing the Rational Speech Act framework with a turn-taking pragmatic system, 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 true interlocutor’s opinion when facing indirect utterances and responses. The model has several novel features: in addition to standard informativeness goals, speaker choices factor in potential divergences of opinions between conversation partners. The listener model further considers multi-turn dialogues rather than isolated utterances: it is able to derive that an utterance like “interesting” can be interpreted positively or negatively depending on preceding discourse. The model, though complex, makes novel, non-trivial qualitative predictions, which are supported by data from three behavioral experiments reported here.