COSPLAY: Concept Set Guided Personalized Dialogue Generation Across Both Party Personas
Introduction
Personalized dialogue systems aim to create more human-like interactions by imbuing agents with specific personas. However, existing models often suffer from "egocentrism," a phenomenon where the agent obsessively broadcasts its own traits while ignoring the user's input and persona. This leads to "stiff" conversations where the agent forces its interests into the dialogue, regardless of the context. For instance, if a user asks about family, an egocentric bot might pivot abruptly to its own hobby of singing, killing the natural flow of conversation.
To address this, we present COSPLAY (COncept Set guided PersonaLized dialogue generation Across both partY personas). COSPLAY shifts the paradigm from individual persona-consistency to a collaborative "team effort," where the agent actively seeks common ground and encourages the partner to express their own persona.
The COSPLAY Framework
The core innovation of COSPLAY is the representation of personas as "Concept Sets" derived from ConceptNet. By moving beyond simple text strings, the model can perform mathematical operations on concepts to find semantic overlaps and related topics.
1. Concept Set Operations
The framework utilizes matrix-based operations to manipulate persona concepts:
- Set Expansion: Finding related concepts (e.g., linking "guitar" to "music").
- Union and Intersection: Identifying shared interests or combining persona traits to broaden the conversational scope.
- Set Distance: Measuring the semantic gap between the dialogue state and the target personas.
2. Model Architecture and Training
COSPLAY is built upon a GPT-2 backbone and follows a two-stage training process: Supervised Fine-tuning (SFT) and Reinforcement Fine-training (RFT).
In the Supervised Fine-tuning phase, the model employs "Concept Attention" and "Concept Copy" mechanisms. These allow the generator to explicitly attend to and incorporate relevant concepts from both the self-persona ($P^S$) and the partner's persona into the response.
In the Reinforcement Fine-training phase, the model is optimized using Policy Gradient methods with two novel rewards:
- Mutual Benefit Reward: This rewards the agent when its response prompts the partner to reveal more of their own persona in subsequent turns. This encourages curiosity and proactive questioning.
- Common Ground Reward: This reward minimizes the geometric distance between the generated dialogue and both personas, forcing the model to navigate topics that are mutually relevant.
Experimental Results
The model was evaluated on the Persona-Chat dataset, including both the Original and the more challenging Revised versions (where personas are rephrased to prevent simple keyword memorization).
Automatic Evaluation
On the Original Persona-Chat benchmark, COSPLAY achieved a Hits@1 of 85.5%, significantly outperforming previous state-of-the-art models like $P^2$ Bot (81.9%). It also secured the highest F1 score (20.16), demonstrating superior language generation and retrieval capabilities. The performance gap was even more significant on the Revised dataset, highlighting the robustness of the concept-based approach over simple word-matching.
Human Evaluation
Crowdsourced human evaluators rated COSPLAY significantly higher than baselines across key metrics:
- Engagement (4.35/5): Users found the conversations more interactive and less repetitive.
- Consistency (4.37/5): The model maintained its persona while naturally integrating the partner's context.
Qualitatively, analysis showed that COSPLAY is notably better at asking natural follow-up questions, fostering a more balanced and human-like dialogue trajectory.
Discussion and Future Work
While COSPLAY represents a significant step toward collaborative AI agents, some limitations remain. The current framework primarily focuses on single-word concepts, which can occasionally lose the nuance of multi-word phrases (e.g., "bird watching"). Additionally, the semantic distances derived from ConceptNet are currently static.
Future research will explore making the concept distance matrix trainable to allow for context-specific semantic relationships. Furthermore, extending the Concept Set logic to handle phrase-level expansions and more complex semantic structures will likely enhance the model's conversational depth.
Conclusion
By treating dialogue as a "mutual benefit" activity rather than a solo performance, COSPLAY effectively mitigates the egocentrism prevalent in modern dialogue agents. The integration of structured concept operations with reinforcement learning provides a powerful framework for generating personalized, engaging, and collaborative conversations.
Summary updated