(Research Center for Language Intelligence of Dalian University of Foreign Languages, Dalian 116044, China)
( School of Software Engineering of Dalian University of Foreign Languages, Dalian 116044, China)
[Objective] This paper focus on the shortage of labeled data resources and the difficulty of distinguishing the weights of sentiment characteristics from the source domain to the target domain in the task of cross-domain sentiment classification.
[Methods] This paper proposed a feature fusion representation method and the attention mechanism based cross-domain product reviews sentiment classification model. This model integrates Bert and cross-domain word vector to generate cross-domain unified feature space, and extracts the importance weight of global and local features through attention mechanism.
[Results] The results of the controlled experiment on public Amazon reviews data set show that the average accuracy of the model reaches the highest value of 95.93% among the control models, 9.33% higher than that of the best control model.
[Limitations] It is necessary to further test the generalization of the model in large-scale multi-domain data sets and explore the contribution of source domain knowledge to target domain review sentiment classification.
[Conclusions] It is feasible to effectively acquire sentiment semantic information by learning fusion features through BiLSTM and attention mechanism. The source fields that are most helpful to the target fields in the controlled experiments are basically the same.