Introduction
Open-domain conversational agents, or chatbots, have seen significant advancements in fluency and response generation. However, two persistent challenges remain: consistency and engagement. Most contemporary models rely on mimicking human-like responses from massive datasets but often lack a stable personality, leading to contradictory statements. Furthermore, they frequently fail to adapt to the specific traits of their interlocutor.
In human conversation, high-quality interaction involves a "theory of mind," where speakers adjust their language based on their perception of the other person. This paper introduces $P^2$ BOT (Persona Perception Bot), a framework designed to explicitly model this mutual persona perception to create more coherent and engaging personalized chatbots.
The $P^2$ BOT Framework
The authors propose a transmitter-receiver architecture that moves beyond simple sequence-to-sequence modeling to incorporate explicit persona awareness.
As shown in the architecture above, the system consists of two primary components:
- The Transmitter: Based on the GPT architecture, this component generates responses by taking the dialogue history and the bot's own predefined persona (a set of profile sentences) as input.
- The Receiver: Based on BERT, this component acts as a "perception" module. It projects dialogue utterances and persona descriptions into a shared latent space to measure how well a specific utterance reflects a given persona.
Training via Mutual Persona Perception
The model is trained using a multi-stage strategy. After initial Supervised Learning via Maximum Likelihood Estimation (MLE), the model undergoes Self-play Reinforcement Learning. In this phase, two bots converse, and the agent is rewarded based on three criteria: language style (fluency), discourse coherence, and—most importantly—Mutual Persona Perception.
This core novelty rewards the agent if the dialogue it produces helps the "Receiver" correctly identify the personas of both speakers. This encourages the bot to act in a way that reveals its own traits while simultaneously "listening" for the other person's traits to maintain flow.
As illustrated in the latent space probing above, the Receiver does not simply look for keyword overlaps. It captures semantic relevance between utterances and traits. For example, it can map the dialogue "I volunteer at the local pool" to the persona trait "I love being in the water," even when the phrasing differs significantly.
Experimental Results
The researchers evaluated $P^2$ BOT using the PERSONA-CHAT dataset, which includes both "Original" and "Revised" (rephrased) persona sentences.
- Automatic Metrics: On the challenging "Revised" set, $P^2$ BOT achieved a Perplexity (ppl) of 18.89, outperforming previous state-of-the-art (SOTA) models like TransferTransfo, which struggled with generalized persona descriptions. An ablation study revealed that the perception reward alone drove a 25% relative improvement in BLEU scores.
- Human Evaluation: In blind tests, humans rated $P^2$ BOT significantly higher than baselines. On a 1-4 scale, it earned an average score of 2.621, compared to approximately 2.0 for the strongest existing models.
Conclusion and Future Work
$P^2$ BOT demonstrates that treating "mutual understanding" as a quantifiable signal for reinforcement learning significantly enhances dialogue quality. While the model occasionally faces challenges with repetitive cycles or redundant questions during self-play, it represents a major step toward chatbots with a true "theory of mind."
Future research may extend the Receiver module to conversational recommendation systems, where a bot could infer a user's preferences through natural interaction to provide personalized content or product suggestions.
Summary updated