Keynote speakers

Attapol RUTHERFORD, PhD

Associate Professor

Department of Linguistics

Chulalongkorn University

Keynote Title

Simulating Discourse with LLM Agents

Bio

Attapol T. Rutherford is a computational linguist at the Faculty of Arts, Chulalongkorn University. His primary research focuses on developing corpora and models for Thai natural language processing and using large language models to study discourse phenomena and computational social science with an emphasis on connecting language use with social constructs.

Abstract

Discourse is perhaps the most complicated level of linguistic analysis because meaning must be interpreted in relation to speakers’ intentions, their interlocutors, and the social context in which an interaction takes place. Studying discourse also presents a practical challenge: the data are often difficult to obtain. This is particularly limiting for corpus and computational approaches, which rely on large amounts of data to reveal linguistic patterns associated with the phenomena under study.

In this talk, we explore the use of large language model (LLM) agents—computational models capable of linguistic and social actions—to simulate discourse and generate corpora for studying specific linguistic phenomena. We test this paradigm by examining persuasion in adversarial multi-party dialogue, where persuasive language can only be understood in relation to the social context and goals of the participants. Rather than treating persuasion simply as a property of individual utterances, we examine persuasive language as social behavior.

We simulate multiple LLM agents playing Coup, a game in which players must persuade and deceive one another to advance their own interests. We present a method for constructing LLM personas and generating dialogue data from interactions among these agents. We then apply text classification and utterance clustering to identify recurring linguistic patterns in persuasive utterances. Our computaional analysis shows that, beyond producing coherent dialogue, LLM agents exhibit several recognizable persuasive strategies: (1) logical persuasion, often realized through conditional constructions; (2) pact formation, commonly expressed through imperatives followed by commitments about future actions; and (3) emotional appeals, realized through affective words and phrasing.

Together, these results demonstrate how LLM-agent simulation can provide a controlled source of interactional data for studying discourse phenomena that are otherwise difficult to observe systematically. More broadly, we argue that LLM agents offer a new experimental paradigm for computational discourse analysis: one in which linguistic behavior can be generated, manipulated, and studied within an explicitly defined social context.

Disclaimer

Site Map