The Pragmatic Parliament: A Framework for Socially-Appropriate Utterance Selection in Artificial Agents

One of the hallmarks of human natural language (NL) interaction is the ability for people to balance a variety of social and communicative goals when choosing how to realize their speech actions. These goals can include pragmatic criteria such as correctness, informativeness, and brevity (i.e., Gricean conversational maxims) or social factors such as politeness. However, there currently does not exist a general algorithmic method to explicitly modulate language generated by artificial agents based on an arbitrary number of pragmatic and social criteria. We propose a novel method to accomplish this task, in which rankings of candidate utterances by different pragmatic or social criteria are fused by use of a voting algorithm. We then give a proof-of-concept demonstration of the application of this method in the context of directive generation for human-robot interaction.