Mining chemical-protein interactions between chemicals and proteins plays vital roles in biomedical tasks, such as knowledge graph, pharmacology, and clinical research. Although chemical-protein interactions can be manually curated from the biomedical literature, the process is difficult and time-consuming. Hence, it is of great value to automatically obtain the chemical-protein interactions from biomedical literature. Recently, the most popular methods are based on the neural network to avoid complex manual processing. However, the performance is usually limited because of the lengthy and complicated sentences. To address this limitation, we propose a novel model, Hierarchical Recurrent Convolutional Neural Network (HRCNN), to learn hidden semantic and syntactic features from sentence sub-sequences effectively. Our approach achieves an F-score of 65.56% on the CHEMPROT corpus and outperforms the state-of-the-art systems. The experimental results demonstrate that our approach can greatly alleviate the defect of existing methods due to the existence of long sentences.