NPC: Personalized Next Profile Crafting Using Previous Role-Play Bots
Abstract: Personalization and role-playing are two important research topics for the LLM community.
However, the exploration in the direction of personalized role-playing especially rare. The primary obstacle in personalized role-playing is the absence of a dataset containing role-playing dialogues with personalized information.
To overcome this obstacle, we introduce a new large-scale personalized role-playing dataset Multi-Bot Tailored Interaction Dataset (MBTI), which includes the entire interaction history from creating bot to deeply engaged conversation between 1238 users and 8477 Bots.
More importantly, we propose a new pipeline called "Next Profile Crafting (NPC)" for crafting role profiles with cross-bot insights to achieve personalization before the conversation. This method is based on the bot persona link among historical bots that user has multi-turn interaction with. We conducted tests using both trained and untrained approaches, as well as open-source and proprietary large language models, highlighting significant disparities in the effectiveness of personalized crafting in the NPC task. Our findings indicate substantial room for improvement in current methodologies.
Paper Type: Long
Research Area: Dialogue and Interactive Systems
Research Area Keywords: Dialogue, evaluation and metric
Languages Studied: English
Submission Number: 5889
Loading