A combined high-quality manual annotation and deep-learning natural language processing study is reported to make accurate name entity recognition (NER) for biomedical literatures. A home-made version of entity annotation guidelines on biomedical literatures was constructed. Our manual annotations have an overall over 92% consistency for all the four entity types such as gene, variant, disease and species with the same publicly available annotated corpora from other experts previously. A total of 400 full biomedical articles from PubMed are annotated based on our home-made entity annotation guidelines. Both a BERT-based large model and a DistilBERT-based simplified model were constructed, trained and optimized for offline and online inference, respectively. The F1-scores of NER of gene, variant, disease and species for the BERT-based model are 97.28%, 93.52%, 92.54% and 95.76%, respectively, while those for the DistilBERT-based model are 95.14%, 86.26%, 91.37% and 89.92%, respectively. The F1 scores of the DistilBERT-based NER model retains 97.8%, 92.2%, 98.7% and 93.9% of those of BERT-based NER for gene, variant, disease and species, respectively. Moreover, the performance for both our BERT-based NER model and DistilBERT-based NER model outperforms that of the state-of-art model,BioBERT, indicating the significance to train an NER model on biomedical-domain literatures jointly with high-quality annotated datasets.
ENCADEAr: ENCADEAmento automático de notícias