Generating Stylistic and Personalized Dialogues for Virtual Agents in Narratives

Virtual agents interact with each other through dialogues in various types of narratives (e.g. films). In this paper, we propose an approach on the basis of DialoGPT pre-trained language model, which explores the impact of dialogue generation with different levels of agents' personalities derived from narrative films based on the Big-Five model, as well as with three different embedding methods. From the experimental results using automatic metrics and human user evaluation, we investigate and analyze the impact of different settings on narrative dialogue generation. We demonstrate that our approach is able to generate dialogues with increased variety that correctly reflect the corresponding target personality.