From Persona to Person: Enhancing the Naturalness with Multiple Discourse Relations Graph Learning in Personalized Dialogue Generation

2025-06-13

AI Summary

This work addresses the challenge of jointly optimizing response naturalness and persona consistency in personalized dialogue generation. Methodologically, we propose MUDI, a Multi-Utterance Dialogue Graph Learning framework: (1) We construct a structured dialogue graph incorporating user persona descriptions to explicitly model heterogeneous utterance-level relations; (2) We design a Dialogue Graph Attention Network (DialogueGAT) encoder coupled with a consistency-aware attention mechanism to jointly encode persona traits, dialogue structure, and generation dynamics; (3) We leverage large language models to assist utterance-relation annotation and graph construction, and integrate coherence-aware attention into decoding. Extensive evaluations on multiple benchmarks demonstrate significant improvements in response naturalness, persona consistency, and discourse coherence—yielding outputs closer to human-level quality. Our core contribution lies in the first deep integration of multi-utterance relational modeling with persona-aware graph neural networks, enabling joint representation of structured semantics and individual characteristics.

Technology Category

• Natural Language Processing: Generation
• Machine Learning: Multimodal Learning
• Data Mining & Knowledge Management: Conversational Systems for Recommendation & Retrieval

Application Category

• User Modeling, Personalization and Recommendation: User modeling and simulation for interactive and conversational systems
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG

Abstract

In dialogue generation, the naturalness of responses is crucial for effective human-machine interaction. Personalized response generation poses even greater challenges, as the responses must remain coherent and consistent with the user's personal traits or persona descriptions. We propose MUDI ($ extbf{Mu}$ltiple $ extbf{Di}$scourse Relations Graph Learning) for personalized dialogue generation. We utilize a Large Language Model to assist in annotating discourse relations and to transform dialogue data into structured dialogue graphs. Our graph encoder, the proposed DialogueGAT model, then captures implicit discourse relations within this structure, along with persona descriptions. During the personalized response generation phase, novel coherence-aware attention strategies are implemented to enhance the decoder's consideration of discourse relations. Our experiments demonstrate significant improvements in the quality of personalized responses, thus resembling human-like dialogue exchanges.

Problem

• Enhancing naturalness in personalized dialogue generation
• Capturing implicit discourse relations with structured graphs
• Improving coherence using novel attention strategies

Innovation

• LLM-assisted discourse relation annotation
• DialogueGAT graph encoder for relations
• Coherence-aware attention strategies for decoding