Abstract: To address the challenges of difficult cross-turn strategy evolution, emotional tone mismatch, and strategy shifts in negotiation dialogue generation, a model integrating Graph Attention Networks and emotional feedback is proposed. In this model, BART is utilised to capture global contextual semantics, and deep structural priors are extracted by constructing a dynamic graph structure containing strategic techniques and dialogue behaviours, combined with multi-head graph attention and adaptive structure-aware pooling mechanisms. To resolve the alignment problem between semantics and strategic logic, structure-aware attention and gating mechanisms are adopted to achieve adaptive information fusion. Furthermore, gradient feedback generated from an emotion auxiliary task is innovatively utilised to globally calibrate the strategy graph representation, ensuring consistency between strategic decisions and the target emotional context. Based on experimental results on the CraigslistBargain dataset, superior performance over existing baseline models is demonstrated across all core metrics, with the BLEU (bilingual evaluation understudy) score reaching 19.74% and the RC-Acc(ratio class prediction accuracy) improving to 54.21% for the negotiation outcome. Finally, the effectiveness of the graph structure modelling and emotion calibration modules in enhancing negotiation logical coherence and emotional adaptability is further validated by ablation studies and human evaluations.

Key words: negotiation dialogue generation, emotion enhancement, BART, graph attention networks, strategy modeling