Deep Learning in Cross-Lingual English-Vietnamese Sentiment Classification

The lack of Vietnamese sentiment resources limits the performance of lexicon-based methods in Vietnamese sentiment classification. However, most of the research is conducted for English language, and English resources are publicly far richer than Vietnamese. Therefore, the main problem we solve is to improve the performance of Vietnamese sentiment classification at document-level by leveraging translated English texts as training data. Vietnamese texts are machine-translated into English texts which are used to extract useful features. Our proposed method is based on Convolution Neural Network model for text classification. We propose to remove insignificant lexical features to speeds up the training process. Furthermore, we propose to replace normal max pooling layer with a static K-max pooling layer at convolutional layer to improve the overall accuracy of the model. Experiment results show our proposed method is able to perform sentiment classification with an accuracy of 84.40% on 25,991 Vietnamese and 25,991 translated English product reviews.