MatSci-YAMZ: Integrating Human and Artificial Intelligence Vocabulary

This paper presents MatSci-YAMZ, a novel implementation of Yet Another Metadata Zoo (YAMZ), a crowdsourced vocabulary application. MatSci-YAMZ is designed to assist vocabulary development for materials science by configuring a large language model (LLM) and provenance tracking into the YAMZ workflow. Implementation of MatSci-YAMZ occurred in three phases: requirements and model selection, development, and fine-tuning. The application utilizes the Gemma3:27b LLM to formulate definitions based upon examples provided with human definitions. Provenance tracking captures the evolution of LLM-generated definition evolution, which occurs in response to user comments. MatSci-YAMZ can improve shared semantics among communities of materials researchers whose terminology may contain nuances or idiosyncratic usages. Improved semantic alignment supports more concise data descriptions and use of metadata, which is foundational to supporting the FAIR principles of findability, accessibility, interoperability, and reusability. Improved provenance tracking also provides greater insight into how researchers engage with LLM-generated definitions as well as the changes that occur. MatSci-YAMZ opens several avenues of research into crowdsourced vocabulary development among materials researchers. Current findings highlight the need for additional user testing and study to better understand human-in-the-loop vocabulary development with existing LLMs.