Large language models for knowledge graph extraction from tables in materials science

Abstract

Research in materials science increasingly harnesses machine learning (ML) models. These models are trained with experimental or theoretical data, the quality of their output hinges on the data's quantity and quality. Improving data quality and accessibility necessitates advanced data management solutions. Today, data are often stored in non-standardized table formats that lack interoperability, accessibility and reusability. To address this issue, we present a semi-automated data ingestion pipeline that transforms R&D tables into knowledge graphs. Utilizing large language models and rule-based feedback loops, our pipeline transforms tabular data into graph structures. The proposed process consists of entity recognition and relationship extraction. It facilitates better data interoperability and accessibility, by streamlining data integration from various sources. The pipeline is integrated into a platform harboring a graph database as well as semantic search capabilities.