A Business Logic-Enhanced Method for Dialogue Policy Learning Based on Large Language Models

• 1、School of Computer Science (National Pilot Software Engineering School), Beijing University of Posts and Telecommunications, Beijing 100876

Abstract：In task-oriented dialogue systems, existing dialogue policy learning methods struggle to meet the robustness and compliance requirements of complex business scenarios due to data limitations, absence of logical constraints, and uncontrollable generation. To address these challenges, this paper proposes a dialogue policy learning method that integrates data augmentation with explicit optimization of business logic. To mitigate data limitations, a hierarchical evol-instruct framework is designed, incorporating breadth-wise evolution, depth-wise evolution, and elimination-based evolution to inject domain-specific rules and logical constraints, thereby enhancing data diversity and complexity.Domain knowledge learning is facilitated through parameter fine-tuning, reducing computational costs while enhancing the model’s adaptability to vertical tasks. Furthermore, a causality-driven business feasibility relation graph is constructed to explicitly define three categories of business logic constraints—"should," "should not," and "can"—thereby optimizing candidate dialogue actions to ensure predictions align with dynamic business rules. Experimental results demonstrate that the proposed method achieves F1 scores of 42.4\% and 81.8\% on the MultiWOZ and SGD datasets, respectively, significantly outperforming baseline models. This work provides an interpretable and controllable technical framework for dialogue policy learning in high-compliance scenarios, offering new insights for advancing the intelligence of task-oriented dialogue systems.

Keywords： Dialogue policy learning Large language model Evol-instruct Parameter fine-tuning Graph augmentation