LLM-Supported Mapping Generation for Semantic Manufacturing Treasure Hunting

In large manufacturing companies, such as Bosch, that operate thousands of production lines with each comprising up to dozens of production machines and other equipment, even simple inventory questions such as of location and quantities of a particular equipment type require non-trivial solutions. Addressing these questions requires to integrate multiple heterogeneous data sets which is time consuming and error prone and demands domain as well as knowledge experts. Knowledge graphs (KGs) are practical for consolidating inventory data by bringing it into the same format and linking inventory items. However, the KG creation and maintenance itself pose challenges as mappings are needed to connect data sets and ontologies. In this work, we address these challenges by exploring LLM-supported and context-enhanced YARRRML mapping generation. Facing large ontologies in the manufacturing domain and token limitations in LLM prompts, we further evaluate ontology reduction methods in our approach. Our work provides a valuable support when creating YARRRML manufacturing mappings as well as supporting data and schema updates. We evaluate our approach both quantitatively against reference mappings created manually by experts and qualitatively with expert feedback.