Rule-based adversarial sample generation for text classification

In Text Classification, modern neural networks have achieved great performance, but simultaneously, it is sensitive to adversarial examples. Existing studies usually use synonym replacement or token insertion strategies to generate adversarial examples. These strategies focus on obtaining semantically similar adversarial examples, but they ignore the richness of generating adversarial examples. To expand the richness of adversarial samples. Here, we propose a simple Rule-based Adversarial sample Generator (RAG) to generate adversarial samples by controlling the size of the perturbation added to the sentence matrix representation. Concretely, we introduce two methods to control the size of the added perturbation, i) Control the number of word replacements in sentences (RAG(R)); ii) Control the size of the offset value added to the sentence matrix representation (RAG(A)). Based on RAG, we will obtain numerous adversarial samples to make the model more robust to adversarial noise, and thereby improving the model’s generalization ability. Compared with the BERT and BiLSTM model baseline, experiments show that our method reduces the error rate by an average of 18% on four standard training datasets. Especially in low-training data scenarios, the overall average accuracy is increased by 12%. Extensive experimental results demonstrate that our method not only achieves excellent classification performance on the standard training datasets, but it still gets prominent performance on few-shot text classification.