Improving Noise Robustness in Text Understanding using Self-Attention Mechanism and GAN

In the process of literary text understanding, the presence of noise may cause the model to misunderstand the text intent. This paper combines the self-attention mechanism with the generative adversarial network (GAN) to learn the dependencies in the text and assign different weights to each word and character in the text during processing, and then introduces GAN to improve robustness. During the training process, the generator and the discriminator are trained adversarially, the generator constantly tries to generate noise text that is closer to the real, and the discriminator constantly improves its ability to identify real text and noise text. The average filtering accuracy of this research method is 95.49%, and the noise robustness improvement rate is between 26.3% and 39.8%. Although this study slightly increases the processing time, it improves the model performance and is effective in improving the robustness of literary texts.