Abstract: Automated contract review faces limitations in dynamically adapting to diverse clauses and accurately executing complex checking rules. To address this, a Context Engineering framework for contract review, termed DRNS-Review (Dynamic Rule Generation with Neuro-Symbolic reasoning), was proposed. The framework first generates review rules in combination with domain knowledge and represents them in a hybrid form of natural language and structured symbols; it then performs neuro-symbolic reasoning, where symbolic matching locates applicable rules and LLM-based semantic execution carries out the reasoning, enabling deep analysis and risk assessment. The approach does not rely on a pre-built knowledge graph, and emphasizes on-the-fly rule generation and application to enhance adaptability and extensibility. On an IT operations and maintenance contract dataset, DRNS-Review improves.

Key words: contract review, large language model, context engineering, dynamic rule generation, neuro-symbolic reasoning, natural language processing

摘要： 自动化合同审查在动态适配多样化合同条款与准确执行复杂审查规则方面存在局限。为此，提出一种基于大语言模型（LLM）动态规则生成与神经符号推理合同审查（DRNS-Review）的“上下文工程（Context Engineering）”创新框架。首先结合领域知识动态生成审查规则，以自然语言与结构化符号的混合形式表达；随后基于神经符号推理，通过符号匹配进行规则定位，并由LLM 语义执行完成推理，最后实现对合同的深度分析与风险评估。该框架不依赖预先构建的知识图谱，强调规则的即时生成与应用以增强适应性与扩展性。在IT 运维服务合同数据集上，相较于直接 LLM 推断与静态人工规则等基线方法，DRNS-Review在风险识别F1值提升了8.3至13.0个百分点，在风险等级准确率（Level Accuracy，LA）提升了8.7至13.3个百分点，不同LLM（如Qwen3、Llama-3）上均取得稳定优势，具有良好的效率性能权衡。实验结果表明，DRNS-Review 通过“上下文工程”实现动态规则生成与神经符号推理，有效提升合同审查的准确性、深度与适应性，为企业级大模型的垂类落地提供可推广的技术路径。

关键词: 合同审查, 大语言模型, 上下文工程, 动态规则生成, 神经符号推理, 自然语言处理