MCP-LSTM Network for Sentence-Level Sentiment Classification

In this paper, an MCP-LSTM network (Multichannel word vector via Convolution and Pooling to concatenate LSTM) is proposed to generate more complete feature representations for sentence sentiment classification. The proposed model combines LSTM and CNN-multichannel network. It uses two set of word vectors, one is a static word vector and the other is a non-static word vector. The two channel word vectors are concatenated. Then the model performs convolution and max pooling on multichannel word vector, convolution layer produces multiple feature maps and max-pooling layer enables the network to focus on the most important feature. Combining the important feature with long sequence information extracted by LSTM, a more complete feature representation of sentences can be obtained. The experimental results demonstrated that the proposed model improves the accuracy of 1.36%, 0.89% and the AUC values of 1.2%, 1.6% on MPQA and MR datasets respectively.