2026 Volume 34 Pages 1115-1123
Research on scientific papers and patents is essential in the field of materials science. However, the considerable volume of literature to be researched renders manual research time-consuming and difficult to quantify. Furthermore, text data are unstructured; hence, they are not fully utilized in information science methods, such as materials informatics. In this study, we propose a method for extracting compound syntheses from scientific papers using a table-filling model based on bidirectional encoder representations from transformers. Although table-filling models perform named entity and relationship extractions for the same text lengths, they are more suitable for long and short text lengths, respectively. The proposed method addresses this contradiction by masking out-of-sentence relationships during training to rebalance the training data. This approach significantly improves relation extraction F1 scores from a baseline of 0.31-0.69 to 0.72-0.85, while maintaining named entity recognition performance. As a result, table-filling model enables highly accurate extraction of compound synthesis methods from scientific literature.