Extracting chemical-protein interactions from biomedical literature via granular attention based recurrent neural networks

The extraction of interactions between chemicals and proteins from biomedical literature is important for many biomedical tasks such as drug discovery and precision medicine. In the existing systems, the methods achieving competitive results are combined of several models or implemented in multi-stage, and they are challenged by high cost because numerous external features are employed. These problems can be avoided by deep learning algorithms, but the performance of the deep learning based models is limited by inadequate exploration of the information. Our goal is to devise a system to improve the performance of the automatic extraction between chemical entities and protein entities from biomedical literature.In this paper, we propose a model based on recurrent neural networks integrating granular attention mechanism. The granular attention can explore the inner information of the context vectors, which are represented in multiple dimensions that play different roles in the extraction of the interactions. Furthermore, we employ Swish activation function in the neural networks for the chemical-protein interactions extraction task for the first time.The proposed method is evaluated on BioCreative VI chemical-protein track test corpus. The experimental results show that this method achieves an F-score of 65.14%, which is 1.04% higher than the state-of-the-art system.The model synthesizing recurrent neural networks and granular attention mechanism, exploring the inner information of the context vectors, can improve the extraction performance without extra hand-crafted features. The experimental results demonstrate that the proposed model is promising for further study on the interaction extraction between chemicals and proteins.