Enhancing Information Extraction from Low-sample Materials Science Literature by Transfer Learning

Extracting information from materials science liter-ature is an important task that can provide data support for data-driven material design methods. This work is not straight-forward, especially in low-sample material domains where the data volume is insufficient to support the accurate training of extraction models. To tackle this problem, we proposes the use of a pre-trained model and a feature-based transfer learning approach to transfer the extraction capabilities learned from large-sample material domains to low-sample domains. To enhance the performance of information extraction, we propose a extraction method based on MatSciBERT and pointer networks. To evaluate the performance of the model, we have annotated datasets across three domains and conducted extensive experiments. The experiments demonstrate that our transfer method can improve extraction models in low-sample domains.