Quantifying the advantage of domain-specific pre-training on named entity recognition tasks in materials science

A bottleneck in efficiently connecting new materials discoveries to established literature has arisen due to an increase in publications. This problem may be addressed by using named entity recognition (NER) to extract structured summary-level data from unstructured materials science text. We compare the performance of four NER models on three materials science datasets. The four models include a bidirectional long short-term memory (BiLSTM) and three transformer models (BERT, SciBERT, and MatBERT) with increasing degrees of domain-specific materials science pre-training. MatBERT improves over the other two BERT_BASE-based models by 1%∼12%, implying that domain-specific pre-training provides measurable advantages. Despite relative architectural simplicity, the BiLSTM model consistently outperforms BERT, perhaps due to its domain-specific pre-trained word embeddings. Furthermore, MatBERT and SciBERT models outperform the original BERT model to a greater extent in the small data limit. MatBERT's higher-quality predictions should accelerate the extraction of structured data from materials science literature.