CAN Complementary Attention Network for Aspect Level Sentiment Classification in Social E Commerce

Customer's reviews play an important role in social e-commerce. It can help merchants establish word-of-mouth and adjust marketing strategies. Because of a large number of customer's reviews, it is difficult to analyze the sentimental polarities of the reviews in time. Aspect level sentiment classification is designed to identify sentimental tendency of the aspect in the sentence. In the previous works, averaged vector was used to represent the original vector usually, but it may lose information when the aspect has multiple words. In order to solve the aforementioned issue, we propose the Complementary Attention Network(CAN), which incorporates the average pooling mechanism, circular convolution method and circular correlation method to complement the information in the averaged vector. We evaluate our model on three datasets: laptop reviews, restaurant reviews and twitter posts. Experiments show that our model is advanced and have achieved excellent performance on multiple datasets. In specifically, we compare different variants of CAN model. The CAN model we proposed achieved better performance on three datasets, especially outperforming A-CAN(attention mechanism) by 3% on laptop reviews.