論文データのMaterials Informaticsに向けた自然言語処理

―超電導材料開発を例に―

抄録

As an emerging style of materials science, we discussed the basic and recent natural language processing technologies, that can be used to collect large experimental dataset for materials informatics. We introduce the classical text-mining for the development of material database, and the approaches to accelerate the automatic data extraction by using recent large language models (LLM). Then we demonstrate how to use the commercial large language models including ChatGPT, by directly asking the LLM how to improve the critical current properties of MgB₂ superconducting wires. By comparing LLM-generated outputs, we analyze word selection and the occurrence of hallucination. Finally, we demonstrate an example to use LLMs effectively, to get inspirations for the development of the best superconducting wire in the history.