Research On Text Sentiment Analysis Based On Adversarial Training

With the vigorous development of applications such as social software and e-commerce platforms,the massive text data existing in these applications contains the public's sentiment attitude towards certain hot events.Analyzing the affective disposition behind these text data has profound implications for public opinion monitoring and strategy formulation.Traditional text sentiment analysis methods provide technical support for mining sentiment attitudes,but the system is less robust.When the text is maliciously added perturbation or destroyed,it will cause certain interference to the sentiment analysis system,resulting in misjudgment of the results.In order to improve the stability and accuracy of the text sentiment analysis system,this paper focuses on the problem of text sentiment classification under the premise of high robustness.The specific research contents are as follows:(1)Aiming at the problem that the text sentiment analysis system is not robust,this paper proposes a text sentiment analysis model that combines the adversarial training and the adversarial dropout.The adversarial training is used to add adversarial perturbation to the text at the input layer to train the adversarial sample,while the adversarial dropout is performed in the hidden layer of the network to dynamically mask the appropriate number of neurons,thereby improving the model robustness and sentiment classification effect.(2)In order to further improve the performance of sentiment classification under the condition of strong robustness,this paper proposes a text sentiment analysis model combining attention mechanism and adversarial training.Attention mechanisms include a global attention mechanism based on sentiment words and a local attention mechanism based on adaptive scale.The former focuses on both sentiment words and the integrity of text information.The latter can adaptively select appropriate scales and capture important local information.Combining the two attention mechanisms with the method of combining adversarial training and adversarial dropout can not only improve the performance of sentiment classification,but also not significant increase the training time.(3)In order to enrich the text feature extractor type,and in order to further improve the performance of sentiment classification,this paper proposes a text sentiment analysis model combining the idea of recurrent self-attention mechanism and adversarial training.A based on recurrent self-attention mechanism is used instead of the recurrent neural network and the convolutional neural network as a text feature extractor,and a residual network structure is used in the module to ensure the performance of the deep network.The module based on the recurrent self-attention mechanism is combined with the method of combining the adversarial training and the adversarial dropout to ensure that the model improves the sentiment classification performance under the condition of strong robustness.