Beyond Whole Dialogue Modeling: Contextual Disentanglement for Conversational Recommendation (DisenCRS)

• Authors: Guojia An, Jie Zou, Jiwei Wei, Chaoning Zhang, Fuming Sun, Yang Yang
• Published: 2025-04-24
• Paper: https://arxiv.org/abs/2504.17427
• Category: Multi-Turn Dialogue / Conversational Recommendation

Background and Motivation

Conversational recommender systems aim to provide personalized recommendations by analyzing and utilizing contextual information related to dialogue. However, existing methods typically model the dialogue context as a whole, neglecting the inherent complexity and entanglement within the dialogue.

Specifically, a dialogue comprises both focus information and background information, which mutually influence each other. Current methods tend to model these two types of information mixedly, leading to misinterpretation of users' actual needs and lower recommendation accuracy.

The Proposed Model: DisenCRS

To address this issue, the paper proposes DisenCRS, a model that introduces contextual disentanglement for conversational recommender systems:

1. Dual disentanglement framework — combines:

• Self-supervised contrastive disentanglement
• Counterfactual inference disentanglement

These jointly distinguish focus information from background information in the dialogue context under unsupervised conditions.

2. Adaptive prompt learning module — automatically selects the most suitable prompt based on the specific dialogue context, fully leveraging the power of large language models (LLMs).

Experimental Results

Experiments on two widely used public datasets demonstrate that DisenCRS significantly outperforms existing conversational recommendation models, achieving superior performance on both:

• Item recommendation
• Response generation

Key Takeaways

• Treating dialogue context as an entangled whole can obscure users' real intent; explicitly separating focus and background signals improves recommendation quality.
• Unsupervised (self-supervised + counterfactual) disentanglement avoids the need for annotated focus/background labels.
• Adaptive prompt selection is a practical mechanism for adapting frozen LLMs to per-dialogue context in recommendation pipelines.

References

• Original paper: Beyond Whole Dialogue Modeling: Contextual Disentanglement for Conversational Recommendation