Abstract: Domain-specific Knowledge Graph (DKG) organize unstructured professional knowledge to enhance downstream tasks. However, existing construction methods often neglect task-specific requirements, leading to redundancy, and rely on costly expert-annotated schemas. To address this, an end-to-end DKG construction framework that eliminates manual schema input was proposed. By decomposing the process into schema induction and knowledge extraction, the framework leverages the LLM "Reasoning-Action-Reflection" paradigm to synchronize schema optimization with constrained extraction. This approach yields task-aligned triples and reusable schemas. Experimental results on datasets from the police and medical fields demonstrate that, without the need for manually predefined precise schema, the proposed method outperforms most baseline methods in terms of Jaccard, recall, and F1 scores, achieving performance close to that of methods constrained by manual schema. This proves the effectiveness of the method in constructing domain knowledge graphs at a low cost.

Key words: Domain-specific Knowledge Graph (DKG), information extraction, Knowledge Graph Construction (KGC), schema constraint, prompt engineering

摘要： 构建领域专有知识图谱（DKG）有助于组织各个专业领域内的大量非结构化知识，更好服务于下游任务。现有知识图谱构建（KGC）方法一方面忽略适配下游任务的需求，造成节点冗余；另一方面，方法依赖专家标注的领域知识模式集作为约束，增加了人工成本。针对以上问题，提出了一种端到端的DKG构建框架。框架在不需要人工模式集输入的情况下，将知识图谱构建任务分解为模式归纳与知识抽取两个部分，遵循推理-行动-反思的大模型（LLM）推理范式，同步进行模式集优化与模式约束下的知识抽取。框架最终输出符合任务需求的知识图谱三元组与可复用于同类任务的模式集。在警务、医疗两个领域数据集上的实验结果表明，在无需人工预定义精确模式集的前提下，本文方法的Jaccard、召回率和 F1分数等指标优于多数基线方法，并取得接近人工模式约束方法的表现，验证了方法在低成本构建领域知识图谱的有效性。

关键词: 领域知识图谱, 信息抽取, 知识图谱构建, 模式约束, 提示工程