📄 中文摘要
大型语言模型（LLM）在语义理解方面表现出卓越的能力，但在需要严谨逻辑的复杂决策任务中，它们往往难以确保结构一致性和推理可靠性。尽管层次分析法（AHP）等经典决策理论提供了系统的理性框架，但其构建严重依赖于劳动密集型的领域专业知识，从而形成了“专家瓶颈”。Doc2AHP模型旨在通过结合LLM的语义理解能力和AHP的结构化决策框架，解决这一挑战。Doc2AHP的核心在于设计了一个创新的语义树表示，该表示能够将非结构化文本信息转化为AHP模型所需的层次结构。具体来说，模型首先利用LLM从文档中识别出决策目标、准则、子准则和备选方案，并构建初始的非结构化知识图谱。随后，通过进一步的LLM推理和专门设计的结构化算法，将这些离散的语义元素组织成一个符合AHP规范的语义树。这个语义树不仅捕获了决策要素之间的层级关系，还编码了它们之间的相对重要性。为了克服LLM在处理数值一致性方面的局限性，Doc2AHP引入了一种迭代优化机制，该机制在语义树构建完成后，对准则和备选方案的权重进行精细调整，以确保其满足AHP的数学一致性要求。此外，模型还开发了一种端到端的评估框架，用于衡量从文档到AHP模型的转换质量，包括语义准确性、结构完整性和决策一致性。实验结果表明，Doc2AHP能够显著减少构建AHP模型所需的人工干预，并在多个领域展现出强大的性能，有效地弥合了LLM的灵活性与传统决策理论的严谨性之间的鸿沟，为自动化复杂决策分析提供了新的途径。
📄 English Summary
Doc2AHP: Inferring Structured Multi-Criteria Decision Models via Semantic Trees with LLMs
Large Language Models (LLMs) demonstrate remarkable proficiency in semantic understanding, yet often struggle to ensure structural consistency and reasoning reliability in complex decision-making tasks demanding rigorous logic. Classical decision theories, such as the Analytic Hierarchy Process (AHP), offer systematic rational frameworks, but their construction heavily relies on labor-intensive domain expertise, creating an “expert bottleneck.” Doc2AHP addresses this challenge by integrating LLM’s semantic comprehension capabilities with AHP’s structured decision framework. At its core, Doc2AHP introduces an innovative semantic tree representation capable of transforming unstructured textual information into the hierarchical structure required by AHP models. Specifically, the model first leverages LLMs to identify decision goals, criteria, sub-criteria, and alternatives from documents, constructing an initial unstructured knowledge graph. Subsequently, through further LLM inference and specially designed structuring algorithms, these discrete semantic elements are organized into a semantic tree compliant with AHP specifications. This semantic tree not only captures the hierarchical relationships between decision elements but also encodes their relative importance. To overcome LLM limitations in numerical consistency, Doc2AHP incorporates an iterative optimization mechanism that refines the weights of criteria and alternatives after semantic tree construction, ensuring they meet AHP's mathematical consistency requirements. Furthermore, an end-to-end evaluation framework is developed to measure the quality of the document-to-AHP model transformation, encompassing semantic accuracy, structural integrity, and decision consistency. Experimental results demonstrate that Doc2AHP significantly reduces the human intervention required for AHP model construction and exhibits strong performance across various domains, effectively bridging the gap between the flexibility of LLMs and the rigor of traditional decision theories, thus offering a new pathway for automating complex decision analysis.