Collaborative attention neural network for multi-domain sentiment classification

Abstract

Multi-domain sentiment classification is a challenging topic in natural language processing, where data from multiple domains are applied to improve the performance of classification. Recently, it has been demonstrated that attention neural networks exhibit powerful performance in this task. In the present study, we propose a collaborative attention neural network (CAN). A self-attention module and domain attention module work together in our approach, where the hidden states generated in the self-attention module are fed into both the domain sub-module and sentiment sub-module in the domain attention module. Compared with other attention neural networks, we use two types of attention modules to conduct the auxiliary and main sentiment classification tasks. The experimental results showed that CAN outperforms other state-of-the-art sentiment classification approaches in terms of the overall accuracy based on both English (Amazon) and Chinese (JD) multi-domain sentiment analysis data sets.