RESIDUAL ATTENTION BI-DIRECTIONAL LONG SHORT-TERM MEMORY FOR VIETNAMESE SENTIMENT CLASSIFICATION
Keywords:Sentiment classification, Neural network, Bi-LSTM, Attention, Residual
Sentiment classification is a problem of assessing and estimating values of people’s opinions, sentiments, and attitudes to products, services, individuals, and organizations. Sentiment analysis helps companies understand their customers for improving marketing strategies in e-commerce, manufacturers decide how to improve their products, or people adjust behavior in their lives. In this paper, we propose a deep network model to classify the reviewed product in Vietnamese. Specifically, we develop a new deep learning model called the Residual Attention Bidirectional-Long Short Term Memory (ReAt-Bi-LSTM) model. First, the residual technique is used in multiple layers Bidirectional Long Short Term Memory (Bi-LSTM), to enhance the model’s capability in learning high-level features from input documents. Second, the attention mechanism is integrated after the last Bi-LSTM layer to assess each word’s contribution to the context vector to the document’s label. Last, the document’s final representation is the combination of the context vector and output of Bi-LSTM. This representation captures both context information from the context vector and sequence information from the Bi- LSTM network. We conducted extensive experiments on four common Vietnamese sentiment datasets. The results show that our proposed model improves the accuracy compared with some baseline methods and one state of the art model for the sentiment classification problem.