Generating Adversarial Examples for Sentiment Classifier of Chinese Sentences

Studies have shown that deep learning models are vulnerable to adversarial examples, which cause incorrect predictions by adding imperceptible perturbations into normal inputs. The same characteristic goes for the sentiment orientation classification models. If the input text for the models contains perturbing information such as typos or special symbols, the inputs could misled the models. The adversarial examples reflect the diversity of text features, and the flaws of the sentiment orientation classification models can be found by adversarial attacks. The adversarial examples can be used to train the sentiment orientation classification models to improve the robustness of the model. This paper proposes to generate adversarial examples of Chinese sentences by replacing one of the characters in the word with similar Chinese characters in a black-box manner. The experimental results show that the adversarial examples generated by this method cost less and visually closer to the normal text.