基于大语言模型与思维链机制的专利功效词自动抽取研究

Automatic Extraction of Patent Effect Terms Based on Large Language Models and Chain-of-Thought Mechanisms

摘要: 专利功效词是构建技术功效矩阵与技术功效图的基础，也是技术情报分析与创新决策的重要语义单元。然而，由于专利文本专业性强、结构复杂且功效表述分散，现有自动抽取方法在准确性与可控性方面仍存在不足。大语言模型在长文本理解方面具有优势，但仅依赖其隐式推理能力，难以稳定识别与核心技术相关的关键功效。针对上述问题，本文将专利功效词抽取重构为引入推理约束的生成式任务，重点探讨思维链机制在该任务中的作用。通过对比无显式推理、通用分步推理以及融入专利分析专家认知逻辑的人工思维链方法，系统分析不同推理引导策略对抽取效果的影响。实验结果表明，引入专家认知逻辑并施加因果约束的思维链设计，能够显著提升抽取质量与稳定性，为技术功效图构建与创新决策支持提供可靠数据基础。

Abstract: Patent effect terms constitute the foundational semantic units for constructing technology effect matrices and technology effect maps, and they play a critical role in technology intelligence analysis and innovation decision-making. However, due to the high level of specialization, complex structure, and dispersed expression of effect descriptions in patent texts, existing automatic extraction methods still exhibit limitations in terms of accuracy and controllability. Although large language models demonstrate strong capabilities in long-text understanding, relying solely on their implicit reasoning abilities makes it difficult to consistently identify key effect terms closely related to core technologies. To address this issue, this study reconstructs the task of patent effect term extraction as a generation-based task guided by explicit reasoning constraints, with a particular focus on the role of the chain-of-thought mechanism. By comparing approaches without explicit reasoning, general step-by-step reasoning, and manually designed chain-of-thought strategies incorporating the cognitive logic of patent analysis experts, this paper systematically examines the impact of different reasoning guidance strategies on extraction performance. Experimental results indicate that a chain-of-thought design integrating expert cognitive logic and causal constraints can significantly improve extraction quality and stability, thereby providing reliable data support for technology effect map construction and innovation decision-making.