Leveraging LLM Reasoning for Knowledge Graph Construction: Capabilities, Methodologies, and Implications

In this paper we introduce a formal approach which employs Reasoning Large Language Models for constructing knowledge graphs (KGs) from historical documents. The method combines named entity recognition techniques with LLM reasoning to align named entities with all their mentions in historic texts and to extract context-aware relationships with supporting evidence. The extracted information is structured to build a knowledge graph, prepared for import, visualization, navigation and validation with the use of graph databases. While the experimental evaluation employs GPT-5 Thinking for LLM reasoning and Neo4j for storing the knowledge graph, the Knowledge Graph construction workflow formalizes the processing pipeline, which can be transcoded and implemented with alternative reasoning LLMs and different graph databases. This approach is validated by analyzing the Dumitru Nistor’s War Diary, a historic document rich in mentions of named entities (i.e. persons, organizations, locations) and historical events related to the first world war. An example for employing this process to generate knowledge graph representations for historical documents, and to generate meaningful visualizations with minimal human oversight is showcased in the paper.