Dynamic Persona Generation through commonsense inference in dialogues

초록

Generating consistent responses reflecting personal information is crucial for building humanlike dialogue systems. However, prior studies have primarily focused on utilizing predefined personas, overlooking scenarios where new facts emerge during a conversation and require the personas to be expanded in real-world applications. To tackle this issue, we propose a novel framework, Dynamic Persona Generation and Selection (DPGS), which automatically generates a speaker's personas by considering previous utterances and selects those containing new information. DPGS reformulates the process of persona generation from previous utterances as a task for performing contextualized commonsense reasoning based on these utterances. Experiments on PersonaChat demonstrate the effectiveness of our framework in both persona generation and response generation.

키워드