From Unstructured Procedural Text to Knowledge Graphs: Comparing LLM-Based Extraction Strategies

Abstract

As experienced workers retire across industrialized economies, organizations risk losing procedural expertise that often remains tacit, undocumented or scattered across unstructured documents. Translating this knowledge into structured, machine-readable representations is difficult to scale, labor-intensive, and prone to inconsistency when done manually. This paper addresses the automated construction of knowledge graphs from natural language procedural descriptions, developing a generic approach for transforming unstructured expert knowledge into structured knowledge graphs that support downstream retrieval and question-answering applications. Three text-to-knowledge-graph approaches were designed, implemented, and systematically evaluated. The first employed a large general-purpose language model (Qwen3-32B) with a single-stage zero-shot extraction prompt, the second applied the same strategy using a smaller base model (Llama2-13B), and the third combined supervised fine-tuning of the smaller model on synthetic extraction data with a decomposed extraction architecture targeting one to two ontology elements per phase. All approaches were evaluated across six procedural descriptions spanning multiple technical domains, with chunk size, model temperature, and ontology detail as configuration parameters. Results were assessed for intrinsic quality and extrinsic fitness for use, measured by question-answering accuracy in a Graph-RAG application. The results demonstrate that extraction strategy is a more decisive factor than model capacity. The fine-tuned model achieved a Question Answering (QA) pass rate of 55.3%, compared to 46.8% for the large model and 33% for the small base model, outperforming the general-purpose model on both intrinsic quality metrics and downstream performance. Average node degree, duplication rate, and ontology simplicity emerge as the strongest predictors of retrieval performance. Shorter ontologies consistently outperformed richer ones, suggesting that ontology design should be driven by the intended downstream application rather than semantic completeness. Over-extraction of procedural steps did not degrade performance but instead improved retrieval robustness by providing additional entry points for graph traversal. The findings offer practical guidance for designing scalable, locally deployable knowledge graph construction pipelines for procedural texts under computational and confidentiality constraints.