Constructing Domain-Specific Knowledge Graphs from Unannotated Technical Texts using LLMs

Building knowledge graphs in specific domains presents significant challenges when domain expertise is limited. The primary obstacles include the lack of annotated datasets and domain-specific models. Although large language models (LLMs) enable flexible extraction from unstructured text, their direct application to technical domains often leads to domain mismatch and inconsistent relation representations. This paper presents a hybrid framework that combines a domain-specific extraction model with LLM-based reasoning to build knowledge graphs from unannotated patent abstracts. Using semiconductor patents as a case study, domain-relevant entities are first identified and then refined through iterative LLM prompting to extract relational triplets. The resulting relations are normalized and integrated into a unified knowledge graph. Experimental results indicate improved textual faithfulness and more coherent relation structures compared to domain-only and LLM-only baselines, demonstrating a practical approach for scalable knowledge graph construction in unannotated technical domains.