Abstract: The text sentiment classification task aims to analyze short text sentences and determine their corresponding sentiment categories.In order to solve the problems of lack of large-scale high-quality corpus dataset and insufficient non-uniform importance extraction of text features in the existing models in sentiment classification,this paper proposes a text sentiment classification method based on large-batch adversarial strategy and enhanced feature extraction.Firstly,the text dataset is input into the pre-trained language model BERT to obtain the corresponding word embedding vector representation,and then the BiLSTM is used to further learn the context dependencies in the sequence.Then,the local attention mechanism is combined with the local receptive field weighting of TextCNN to enhance the feature extraction ability.Finally,the output of BiLSTM and the output of TextCNN are spliced to obtain the deep feature fusion of the two spaces,which are handed over to the classifier for the judgment of sentiment classification.In the whole training process,a large-batch adversarial strategy is adopted,and adversarial perturbations are added to the word embedding space and multiple iterations are carried out to improve the robustness of the model.Experimental results on multiple datasets verify the effectiveness of the proposed model.

Key words: Short text,Sentiment classification,Adversarial strategy,Feature extraction,Word embeddings