Automatically extracting composition, process, and property data from concrete papers with an LLM pipeline to build a large-scale DB

1/ AI utilization method: large language model (LLM)-based literature information extraction pipeline
2/ Applied task: extracting composition, process, and property attributes from concrete and blended cement papers to build a structured database
3/ Reported results: maximum F1 0.98, about 9,000 records and more than 100 attributes extracted within 1 hour
4/ This study reported the development of a large language model (LLM) pipeline that automatically extracts and structures material data related to concrete and blended cement from unstructured paper text.
5/ The target task was to extract composition, process, and property attributes from papers and create a database for materials informatics.
6/ Reported achievements include a maximum F1 of 0.98, extraction of about 9,000 records within 1 hour, extraction of more than 100 attributes, and the construction of a large open database based on a corpus selected from more than 27,000 papers.
7/ From a materials informatics perspective, this is a representative use case of converting experimental results into standardized records through literature mining, and it has direct significance for transforming existing literature assets into datasets.
8/ However, the figures presented in the article are results for a specific corpus and task, so they cannot be taken as automatically generalizable to other material systems or other literature styles.
9/ In addition, extraction performance can vary depending on the evaluation criteria and the quality of the ground-truth annotations. [1]