Forecasting outcomes in mixed-motive negotiations requires integrating explicit linguistic cues with latent strategic constraints, such as budgets and alternatives. Existing computational models often fail to adapt to varying task structures and may not adequately account for distributive considerations present in historical training data. This study proposes a unified framework to adaptively fuse semantic and strategic signals while incorporating reflective modeling of utility disparities. We introduce the Semantic-Temporal Graph Fusion Network (ST-GFN), a dual-stream architecture that processes textual dialogue with transformer encoders and economic states with Graph Attention Networks, connected via a dynamic gated fusion mechanism. Evaluated on contrasting benchmarks, the linguistically oriented DealOrNoDeal and the strategy-oriented CaSiNo, ST-GFN exhibits strong adaptability. The model dynamically adjusts modality weighting, emphasizing linguistic cues in free-form settings (z ~ 0.97) and increasing reliance on strategic constraints in structured tasks (z ~ 0.73). A fairness-regularized composite loss is incorporated to penalize deviations from ground-truth utility gaps. Results demonstrate a 43.8% reduction in Inequality Discrepancy in high-disparity environments with minimal impact on accuracy, alongside improved performance in high-variance domains. These findings suggest that reflective regularization can enhance both predictive reliability and equitable representation in negotiation forecasting, supporting the design of transparent Group Decision and Negotiation Support Systems (GDNSS).

翻译：预测混合动机谈判中的结果需要整合显性语言线索与隐性策略约束（如预算和替代方案）。现有计算模型常难以适应多变的任务结构，且未能充分考量历史训练数据中的分配性因素。本研究提出统一框架以自适应融合语义与策略信号，同时引入对效用差异的反思性建模。我们提出了语义-时序图融合网络（ST-GFN），一种双流架构：采用Transformer编码器处理文本对话，结合图注意力网络建模经济状态，两者通过动态门控融合机制连接。在语言导向的DealOrNoDeal基准与策略导向的CaSiNo基准的对比评估中，ST-GFN展现出强适应能力。模型动态调整模态权重，在自由对话场景（z≈0.97）中侧重语言线索，在结构化任务（z≈0.73）中增加对策略约束的依赖。通过引入公平正则化复合损失函数，对偏离真实效用差距的预测进行惩罚。结果表明，在高不平等环境中，不等式差异降低了43.8%且精度损失极小，同时在方差较大的领域性能提升。这些发现表明，反思性正则化可同时增强谈判预测的可靠性与公平表示，为设计透明的群体决策与谈判支持系统（GDNSS）提供依据。