KNOWLEDGE GRAPH CONSTRUCTION FROM MATERIALS SCIENCE LITERATURE USING LARGE LANGUAGE MODELS AND ADVANCED DATA PREPROCESSING

Abstract

In this work, we present a pipeline for automated knowledge graph construction from materials science literature using large language models (LLMs). The proposed method performs entity and relationship extraction guided by a data model based on the logic of the Elementary Multiperspective Material Ontology (EMMO), structuring the output into a machine-interpretable graph format. The pipeline integrates several key components, including prompt-based extraction, a hierarchical chunking strategy that leverages document structure and section headers, and post-processing steps such as normalization, LLM-assisted deduplication, and alignment of node identifiers. A central focus of this study is the evaluation of different chunking strategies. Specifically we compare fixed-size splitting with a hierarchical chunking approach that incorporates document structure and header information. Our results show that hierarchical chunking consistently outperforms fixed-size chunking across both entity and relationship extraction tasks, achieving higher precision, recall, and F1 scores through more context-aware segmentation. Extracted entities and relationships are aligned with a curated ground truth dataset through manual verification to ensure semantic correctness. Overall, these findings indicate that LLMs, when combined with domain-specific ontological guidance and well-designed pre-and post-processing, can effectively extract high quality knowledge graphs from complex materials science literature. This benefits materials scientists and researchers by reducing manual curation effort and accelerating data-driven materials discovery.