Dual-Channel Domain Adaptation Model

Abstract

Document-level cross-domain sentiment analysis aims to leverage useful information in the source domain to help infer document-level sentiment on the target domain. The existing cross-domain sentiment analysis methods neglect complex syntactic structure and diversified semantic information of document text in different domains. Therefore, we proposed a novel dual-channel domain adaptation model (DCDA) for document-level cross-domain sentiment analysis. It consists of feature extraction module and domain adaptation module. The dual-channel feature extraction module adopts hierarchical attention structure to extract context channel features at word level and sentence level. In addition, different attention strategies are implemented at different levels, which enables accurate assigning of the attention weight. GAT is used to extract syntactic channel characteristics of documents. We adopt adversarial mutual learning in the domain adaptation module. It learns the domain-invariant features by using adversarial network learning, and makes full use of the information of the target domain to improve the classification effect by mutual learning. Experiments on multiple public datasets demonstrate the effectiveness of DCDA.