Densely Connected Bidirectional LSTM with Max-Pooling of CNN Network for Text Classification

Text classification is a fundamental task in natural language processing (NLP). Context semantics can greatly improve the accuracy of text classification tasks. Although there are some popular methods in obtaining semantics, current context semantic analysis techniques, due to limited accuracy, are still a great bottleneck for text classification. This paper introduces a novel model, the densely connected Bidirectional LSTM with Max-pooling of CNN network (Dense-BiLSTM-MP), which greatly enhances the context of semantic information. In this model, a densely connected bidirectional long short-term memory (BiLSTM) model, as well as multiple max-pooling layers of convolutional network, are applied to obtain an increasingly enhanced assessment of context, and extract the key features, respectively. Experiments were conducted on four public datasets: YELP, 20NewsGroup, THUNews and AG. The experimental results show that the proposed model outperforms state of the art methods on several datasets. Furthermore, discussions on the Dense-BiLSTM-MP model’s performance in short texts and long texts were given, respectively.