Method of the E-commerce Review Sentiment Classification Based on BiGRU-multi-head Attention
In view of the problems in e-commerce review sentiment classification such as numerous model's parameters and parameters random initialization,a combined model based on bidirectional gated recurrent neural network(BiGRU)and multi-head attention is proposed.First,BiGRU model is used to extract more comprehensive semantic information from the positive and negative bidirections.After that,the attention information is extracted by parallel computation in several attention heads.Final-ly,the feature vectors combining the two aspects of information is input into the classifier,and this model is constructed.The param-eters of the model are debugged.The accuracy rate and cross entropy are used as evaluation criteria to compare the sentiment classifi-cation effect of four mainstream deep learning models on the e-commerce review data.These models to be compared include bidirec-tional long short-term memory neural network(BiLSTM),BiGRU,BiLSTM and single-head attention,BiGRU and single-head at-tention,BiLSTM and multi-head attention,and BiGRU and multi-head attention.The results show that on the same datasets,accu-rary of the proposed model has reached 89.91%,and has better performance in accuracy rate and cross entropy compared with oth-ers.
BiGRUmulti-head attentionsentiment classificatione-commerce review data
国家科技期刊平台
NSTL
万方数据
