DE-CO: A Two-Step Spelling Correction Model for Combating Adversarial Typos
Authors:
- 1 December 2020
- conference paper
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 554-561
Abstract
The robustness of text classifier based on the deep neural networks can be improved by correcting adversarial spelling mistakes. The main challenge brought by these mistakes is that it is difficult for the text classifier to recognize the words in the text correctly. Existing works to deal with these mistakes can be divided into two types: optimization of training data and reconstruction of text classification model. However, these works are not suitable for the text classification model that has been deployed, as retraining or reconstruction is always involved. To address the above problem, we propose a two-step spelling correction model, which consists of a misspelled word detector and a misspelled word correcter, referred to as DE-CO. Specifically, we use a detector to recognize misspelled words in the text, use a corrector to correct the misspelled words, and then feed the correction results into the downstream classifier. In this way, without reconstruction or retraining, the normal recognition of words by the text classifier can be guaranteed. We evaluate DE-CO on adversarial examples generated from the Stanford Sentiment Treebank (SST). The classification accuracy of the downstream classifier is improved in a variety of attack scenarios, which demonstrates that DE-CO improves the robustness of the text classifier.Keywords
This publication has 14 references indexed in Scilit:
- TextBugger: Generating Adversarial Text Against Real-world Applications2019
- Deep Text Classification Can be Fooled2018
- Interpretable Adversarial Perturbation in Input Embedding Space for Text2018
- Black-Box Generation of Adversarial Text Sequences to Evade Deep Learning Classifiers2018
- Towards Robust Neural Machine Translation2018
- HotFlip: White-Box Adversarial Examples for Text Classification2018
- Breaking NLI Systems with Sentences that Require Simple Lexical Inferences2018
- Adversarial Examples for Evaluating Reading Comprehension Systems2017
- Crafting adversarial input sequences for recurrent neural networks2016
- End-to-end Sequence Labeling via Bi-directional LSTM-CNNs-CRF2016