You Impress Me: Dialogue Generation via Mutual Persona Perception

Qian Liu, Yihong Chen, Bei Chen, Jian-Guang Lou, Zixuan Chen, Bin Zhou, Dongmei Zhang

Abstract

Open-domain dialogue systems often struggle with consistency and genericness. This paper introduces P² BOT (Persona Perception Bot), a framework designed for "mutual persona perception." By employing a transmitter-receiver architecture and reinforcement learning, P² BOT optimizes for conversations that actively reveal and reinforce the unique personalities of both speakers. Our approach achieves state-of-the-art results on the PERSONA-CHAT dataset and significantly improves human-rated consistency and engagement.

Introduction

A persistent challenge in open-domain dialogue is the "generic response" problem, where bots provide safe but uninformative answers. Furthermore, these systems often fail to maintain a consistent persona throughout a conversation. While previous research has grounded dialogue in persona profiles (sets of descriptive sentences), these models primarily focus on word-level mimicry rather than a deeper understanding of the interlocutor's identity.

Example of a persona-based dialogue where speakers reveal traits through conversation.

The authors argue that effective human communication involves a shared mental state where participants gradually "perceive" each other's personas. P² BOT addresses this by treating understanding as a mathematical reward signal, ensuring that the dialogue serves as a medium for identity exchange.

The P² BOT Framework

The P² BOT architecture utilizes a transmitter-receiver framework to facilitate mutual understanding.

The P² BOT framework featuring a transmitter-receiver architecture for mutual persona perception.
  1. The Transmitter: Based on a pre-trained GPT model, this component generates responses by taking the dialogue history and the bot's own persona as input.
  2. The Receiver: Based on BERT, this component acts as the "mind" of the interlocutor. It analyzes the dialogue to "guess" the speaker's persona, projecting utterances and persona descriptions into a shared latent space to measure alignment.

Training and Reward Shaping

The model is first trained using supervised learning to predict human responses. It is then fine-tuned through Self-Play, where a learnable agent interacts with a frozen "user" agent. The agent is rewarded via Reinforcement Learning (RL) based on three signals:

  • Language Style: Evaluated by a language model to ensure human-like speech.
  • Discourse Coherence: Ensures the response logically follows the previous turn.
  • Mutual Persona Perception: A novel reward given if the agent's responses help the Receiver correctly identify its persona, and if the agent asks questions that help identify the partner's persona.

Experimental Evaluation

The system was evaluated on the PERSONA-CHAT dataset, including a "Revised" mode where persona descriptions are rephrased to test underlying trait understanding rather than simple keyword matching.

Key Results

  • Automatic Metrics: P² BOT achieved state-of-the-art results in the Revised mode, reaching a Perplexity (ppl) of 15.12 and an F1 score of 19.77, a significant improvement over previous models like TransferTransfo.
  • Human Evaluation: In blind tests, human judges rated P² BOT significantly higher (2.62/4.0) than competitors (1.95/4.0) in terms of consistency and engagingness.
  • Persona Probing: The Receiver component proved highly effective, achieving over 93% accuracy in identifying the correct persona from a dialogue in the original mode.

Conclusion

P² BOT represents a shift from simple mimicry to intentional communication. By framing persona perception as a long-term goal, the model generates more engaging and consistent conversations. Future work will focus on reducing repetition and applying the "Receiver" concept to conversational recommendation systems.

Summary updated

BibTeX
@inproceedings{liu2020you,
  author = {Qian Liu and Yihong Chen and Bei Chen and Jian-Guang Lou and Zixuan Chen and Bin Zhou and Dongmei Zhang},
  title = {{You Impress Me: Dialogue Generation via Mutual Persona Perception}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2020},
  doi = {10.18653/v1/2020.acl-main.131},
}
Figure