MORPHEUS: Modeling Role from Personalized Dialogue History by Exploring and Utilizing Latent Space

Yihong Tang, Bo Wang, Dongming Zhao, Jinxiaojia Jinxiaojia, Zhangjijun Zhangjijun, Ruifang He, Yuexian Hou

Abstract

Personalized Dialogue Generation (PDG) aims to create coherent responses according to roles or personas. Traditional PDG relies on external role data, which can be scarce and raise privacy concerns. Approaches address these issues by extracting role information from dialogue history, which often fail to generically model roles in continuous space. To overcome these limitations, we introduce a novel framework Models Roles from Personalized Dialogue History by Exploring and Utilizing Latent Space (MORPHEUS) through a three-stage training process. Specifically, we create a persona codebook to represent roles in latent space compactly, and this codebook is used to construct a posterior distribution of role information. This method enables the model to generalize across roles, allowing the generation of personalized dialogues even for unseen roles. Experiments on both Chinese and English datasets demonstrate that MORPHEUS enhances the extraction of role information, and improves response generation without external role data. Additionally, MORPHEUS can be considered an efficient fine-tuning for large language models.

Introduction

Personalized Dialogue Generation (PDG) is essential for developing chatbots that exhibit consistent character traits and individual styles. Traditionally, this task has relied on explicit persona descriptions—short text snippets like "I am a teacher who loves jazz." However, such data is often unavailable in real-world applications and poses significant privacy risks. While history-based models attempt to infer personas from past conversations, they frequently struggle with role isolation—treating every user as a unique entity without recognizing shared traits—and poor generalization to "unseen" roles not present in the training set.

MORPHEUS addresses these challenges by mapping dialogue history into a structured "persona codebook" within a shared latent space. This allows the model to understand underlying traits rather than just memorizing specific user histories, enabling consistent and diverse response generation even for novel personas.

Methodology

The MORPHEUS framework (Models Roles from Personalized Dialogue History by Exploring and Utilizing Latent Space) employs a three-stage training pipeline designed to capture and utilize latent role information effectively.

The MORPHEUS framework architecture, illustrating (a) Role Awareness, (b) Persona Codebook Initialization via EM, and (c) Joint Training with VQ-VAE and contrastive learning.
  1. Role Awareness: The model is first fine-tuned to generate responses using both dialogue history and explicit persona data. This stage teaches the decoder to perceive the relationship between specific traits and linguistic patterns.
  2. Persona Codebook (PC) Initialization: A key innovation of MORPHEUS is the initialization of a discrete codebook of $N$ latent representations. Using the Expectation-Maximization (EM) algorithm, the authors cluster existing persona data to provide a global perspective on the variety of possible roles, ensuring a more stable training process compared to random initialization.
  3. Joint Training: The model learns to predict the optimal "code" from the codebook for a given dialogue history. This stage utilizes a Vector Quantized Variational Autoencoder (VQ-VAE) approach to maintain a compact codebook and a contrastive learning loss to ensure that the learned representations are diverse and distinct.

Notably, MORPHEUS can function as a Parameter-Efficient Fine-Tuning (PEFT) method. By freezing a large language model (LLM) and only training parameters related to the persona codebook, it offers a computationally efficient path to personalization.

Experimental Setup

The researchers evaluated MORPHEUS on two major benchmarks:

  • ConvAI2: A standard English dataset for personalized dialogue based on the PersonaChat corpus.
  • Baidu PersonaChat: A large-scale Chinese dataset mirroring the structure of ConvAI2.

During training, explicit persona descriptions were provided, but these were masked during inference to simulate real-world scenarios where only dialogue history is available.

Results and Discussion

MORPHEUS consistently outperformed state-of-the-art (SOTA) baselines, including MSP and CLV, across multiple evaluation metrics.

Automatic evaluation results on ConvAI2 and Baidu PersonaChat datasets. MORPHEUS outperforms state-of-the-art baselines across most metrics.

Key findings include:

  • Coherence and Consistency: On the ConvAI2 dataset, MORPHEUS achieved a BLEU-1 score of 12.67 and a Consistency score (P-Co) of 11.64, significantly higher than the previous SOTA, CLV (11.16 and 9.38, respectively).
  • Diversity: By utilizing contrastive learning, MORPHEUS avoids "mode collapse"—a common issue in latent variable models where responses become repetitive and generic. It maintained high Dist-1 and Dist-2 scores.
  • Human Evaluation: Blind tests confirmed that human evaluators preferred MORPHEUS for its readability, diversity, consistency, and coherence. It achieved a coherence score of 0.85, approaching the ground-truth score of 0.91.

The results suggest that dialogue history is a "dense" signal. Even without an explicit biography, the nuances of how a user discusses their hobbies or profession provide sufficient information for the model to map them to the correct latent "code" in the persona codebook.

Conclusion

MORPHEUS represents a significant advancement in personalized dialogue generation by moving away from a reliance on explicit persona data. By exploring and utilizing a latent persona space, it achieves superior generalization and consistency. Future work will explore expanding the linguistic scope beyond English and Chinese, optimizing dynamic codebook structures, and addressing the ethical implications of generating highly personalized AI personas.

Summary updated

Abstract

Personalized Dialogue Generation (PDG) aims to create coherent responses according to roles or personas. Traditional PDG relies on external role data, which can be scarce and raise privacy concerns. Approaches address these issues by extracting role information from dialogue history, which often fail to generically model roles in continuous space. To overcome these limitations, we introduce a novel framework Models Roles from Personalized Dialogue History by Exploring and Utilizing Latent Space (MORPHEUS) through a three-stage training process. Specifically, we create a persona codebook to represent roles in latent space compactly, and this codebook is used to construct a posterior distribution of role information. This method enables the model to generalize across roles, allowing the generation of personalized dialogues even for unseen roles. Experiments on both Chinese and English datasets demonstrate that MORPHEUS enhances the extraction of role information, and improves response generation without external role data. Additionally, MORPHEUS can be considered an efficient fine-tuning for large language models.
BibTeX
@inproceedings{tang2024morpheus,
  author = {Yihong Tang and Bo Wang and Dongming Zhao and Jinxiaojia Jinxiaojia and Zhangjijun Zhangjijun and Ruifang He and Yuexian Hou},
  title = {{MORPHEUS: Modeling Role from Personalized Dialogue History by Exploring and Utilizing Latent Space}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2024},
  url = {https://openreview.net/forum?id=en7sNsRho6},
}
Figure