Ontology-Driven Control of Conversational LLMs for AI Problem Formulation

(2pm, room C005)

LLM-based conversational agents produce fluent text but lack an explicit conversational plan, drifting off topic, and, when small, collapsing entirely over long conversations. We investigate whether ontologies can supply the missing control and structure. We present BOLIDE, a dialogue system which uses a conversational strategy ontology to guide the LLM agent using dialogue acts and a memory ontology to record the user’s answers. We find that control eliminates the catastrophic degeneration of the smaller models and improves the completeness of the recorded information. Control also reduces generation artifacts, off-topic rambling, and enforces the impartial interviewer role. Our system allows smaller locally run models be used for the interview task by separating the structure and memory of the interview from the model capability.