Joint Dilated Convolution and Self-attention for Cross-domain Sentiment Analysis

Abstract Cross-domain sentiment analysis aims to use source domain resources or models serve the tasks in target domain, and it can effectively alleviate problem of insufficient tagged data domain. In this paper, we investigate deep transfer learning for text analysis, focus on customer reviews different domains. Given shortcomings conventional neural network, such as convolutional network (CNN) cannot obtain long-term dependence between features, recurrent (RNN) achieve parallel computing, propose a novel which joint dilated convolution self-attention (ADC). The ADC takes residual its basic framework, capture more distant information by introducing dilation rate convolution. Self-attention is used make pay attention important information. First, pre-trained with weight pre-training frozen. Then, model fine-tuned method gradual unfreezing, updated through Finally, be utilized analysis. By using review datasets, carry out extensive experiments demonstrate that has superior performance better cross-domain