Simulating discourse with LLM agents

Discourse phenomena in linguistics have been an ambition and interest for computational linguists for decades, but the available computational models were underpowered for analyzing the linguistic phenomena at that level. These phenomena, such as discourse coherence, positive framing, etc., are often complex, as the meaning goes beyond the surface lexical or syntactic analysis. Their meaning arises from the social context and from speakers’ and hearers’ intentions in a particular interaction. Understanding discourse phenomena matters to computational linguists because it can help us build effective dialogue systems that connect world knowledge, social context, and meaning. Medical and tutoring dialogue systems, for example, need to interpret more than the literal content of individual utterances (Grenander et al., 2021; Xu et al., 2023).