TTI-COIN at BioCreative VII Track 2 Fully neural NER, linking, and indexing models

We built neural models that extract chemical entities from full papers, link them to database entries, and select key terms from the extracted terms. All our models adopt pretrained BERT language models, i.e., SciBERT and PubMedBERT, enabling the training of high-performance models from a small amount of data. Keywords–—linking; indexing; neural network; representation learning; TF-IDF