Transformer-Based BiLSTM for Aspect-Level Sentiment Classification

In order to further improve the effect of sentiment classification of multi-sentiment sentences, a hybrid model based on BiLSTM and aspect Transformer is proposed. First, BiLSTM is used to extract sentence context features, and then the obtained features are trained as multi-aspect Transformer modules, each of which is independent of each other. During the training, the parameters of each Transformer module are constantly adjusted to accurately refining the sentimental polarity of the sentence. Experimental results on SemEval data set show that the proposed method can effectively improve the accuracy of sentiment classification and optimize the performance of sentiment classification.