Hierarchical Neural Network with Serial Attention Mechanism for Review Sentiment Classifification

Hierarchical neural networks have achieved success in review sentiment classification. Recently, some works achieved improvements by integrating user and product information to build a review representation. However, it neglects comprehensive underlying relationship between meaningful information and contextual sentences in a review since the full meaning of review is made up of both several independent sentences and a serial of continuous sentences. In this paper, we propose a hierarchical neural network with the serial attention mechanism in sentence level to extract the association of the sentences to the meaningful information, which is useful for generating a more accurate review representation. Firstly, it learns two representations for each sentence of a review from two individual hierarchical networks with user attention or with product attention. Secondly, based on the two representations of a sentence, it computes two representations for each review with individual serial attention mechanism from user perspective and product perspective. Finally, a fusion gating mechanism is used to select the two review representations from user perspective and product perspective so as to determine the final representation of the current review. Experimental results on three datasets validate the obvious effectiveness of our model.