Biomedical Named Entity Recognition via A Hybrid Neural Network Model

Abstract

Biomedical named entity recognition (BNER) is one of the primary tasks of analyzing and mining biomedical resources. Recently, major neural network models such as convolutional neural network (CNN) and long short-term memory network (LSTM) have shown better performance than conventional methods in BNER. They do not rely on any feature engineering. However, CNN and LSTM have their own advantages for BNER. Combining them effectively based on these advantages may improve the performance of the BNER model. We present a hybrid model by combining CNN with bidirectional LSTM (BLSTM). The hybrid dilation convolution structure and the attention mechanism are adopted to improve the two networks respectively. They can obtain a wider range of contextual information and increase the weight of significant information. First, we convert the original data into the word vectors. Then the hybrid dilation CNN and the attention-based BLSTM are respectively used to capture features. Next, the output features are integrated by the merge operation. Finally, the conditional random field (CRF) layer is adopted to parse the tags. Compared with some previous single-task BNER methods, our hybrid model obtains the reasonable performance (F1-score:73.72%) on the JNLPBA corpus without any feature engineering, only with public general pretrained word embeddings and simple architecture. To the best of our knowledge, we are the first to adopt the hybrid structure which combines CNN with BLSTM in BNER.