Multi-way matching based fine-grained sentiment analysis for user reviews

While sentiment analysis has been widely used in public opinion to explore tendency of users for a target product from large online review data, less work focus on aspect-level or fine-grained sentiment analysis in which the polarity of not only the aspect of a target object but also the attribute of that given aspect should be determinated. Recent work regards aspect-level sentiment analysis as two separate tasks, i.e., aspect classification and sentiment analysis, and this pipeline method leads to error propagation. To address this issue, this paper proposes an improved multi-way matching deep neural network model for fine-grained sentiment analysis, which jointly models the two tasks in one phase and improves current attention by directly capturing past attention in the multi-round alignment architecture, so as to prevent error propagation and attention deficiency problems. Experimental results on fine-grained sentiment analysis data sets of catering industry indicate that the F1 score of this model in actual test set reaches 0.7302 and EM score 87.1973, which are higher than baseline DocRNN model by 3.8% and 0.88% in F1 and EM, and are higher than SVM by 15.4% and 25.6%, which verified that our model could effectively predict fine-grained sentiment and have better generalization performance.