基于信息抽取的基因编辑技术领域大模型偏见评估研究

Bias Assessment of Large Language Models in Gene Editing Technology through Information Extraction

摘要: 基因编辑技术是一项兼具重大应用潜力与生物安全风险的前沿技术。当前，大语言模型在其相关科研、传播中的应用日益广泛。如果大模型对基因编辑技术存在认知偏见，将可能引发生物安全风险。为系统评估大模型对基因编辑技术的认知偏见，本研究构建了一种基于事件要素的信息抽取分析框架。该框架聚焦“人物、组织、技术、对象、效果、发表期刊”六类核心事件要素，采用“Basic-Rethink-Multi-Query”的级联评估流程，并在gpt-3.5-turbo与gpt-4-turbo两类模型及多种提示策略下进行实验验证。研究结果显示，基础信息抽取环节存在显著的结构性偏见，表现为对静态实体(如技术、组织、发表期刊)的识别效果较好，而对动态要素(如人物、对象、效果)的识别能力较弱。在引入反思机制与多轮问答策略后，各要素的识别均衡性与整体性能均得到显著提升，但“对象”等特定要素的识别仍存在滞后，提示模型在领域语义理解上存在盲区。本研究通过信息抽取方法有效识别并缓解了大模型在基因编辑技术领域的认知偏见，为发展可信赖的信息处理技术、支持生物安全治理提供了方法依据与实证参考。

Abstract: Gene editing technology represents a rapidly advancing field with substantial application potential and significant biosecurity risks. Large language models (LLMs) are increasingly used for scientific research and communication in this domain. If such models exhibit cognitive bias in gene editing technology, they may exacerbate biosecurity risks. To systematically assess such bias, we propose an event-centric information extraction framework. The framework targets six core event elements—Person, Organization, Technology, Object, Effect, and Publish—and employs a cascaded evaluation pipeline (“Basic-Rethink-Multi-Query”). We evaluate gpt-3.5-turbo and gpt-4-turbo under multiple prompting strategies. Results reveal significant structural bias in the basic extraction stage: the models perform better on relatively static entities (Technology, Organization, Publish) compared to dynamic elements (Person, Object, Effect). Incorporating Rethink mechanisms and Multi-Query strategy substantially improves both inter-category balance and overall extraction performance. However, extraction of certain elements (e.g., Object) remains comparatively weak, indicating gaps in domain-specific semantic understanding. This study applies information extraction methods to identify and mitigate LLM cognitive bias in the domain of gene editing technology, thereby providing a methodological foundation and empirical evidence for trustworthy information processing technologies and biosecurity governance.