Flat and Nested Protein Name Recognition Based on BioBERT and Biaffine Decoder

With the rapid growth in the volume of literature in the biomedical field, it has become increasingly important to extract key information from it in an automated manner. Protein name recognition is of great research importance as an important pre-task for the automatic extraction of protein-protein interactions. Many existing recognition methods focus on flat protein name entities and have difficulty in handling nested entities. Existing nested entity recognition methods have some limitations, such as poor robustness due to overconfidence and limited word embedding characterization ability. To address these problems, a protein name recognition method based on BioBERT and biaffine decoder are proposed in this work. BioBERT pre-trained weights constructed on a large-scale biomedical corpus are used to improve the characterization ability of the embedding layer and biaffine decoder is used to achieve the representation of nested entities. Moreover, a smoothing strategy is applied to improve the biaffine decoder’s selection of entity boundaries, which can improve the accuracy of entity recognition. Experiments on flat and nested named entity recognition datasets containing a large number of protein name entities show that the proposed method can achieve better performance than previous state-of-the-art methods (78.59% and 82.56% in F1-scores on JNLPBA and GENIA datasets, respectively).