Research map / Adversarial robustness and security / Certified and empirical robustness
Unifying Model Explainability and Robustness for Joint Text Classification and Rationale Extraction
Dongfang Li, Baotian Hu, Qingcai Chen, Tujie Xu, Jingcong Tao, Yunan Zhang
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use AT-BMC with BERT-base instead of the strongest listed baseline | text classification and rationale extraction · Movie Reviews | Token F1: 46.3 -> 50.6 | Improved |
| Use AT-BMC with BERT-base instead of the strongest listed baseline | text classification and rationale extraction · MultiRC | Token F1: 54.0 -> 57.3 | Improved |
| Use AT-BMC with BERT-base instead of vanilla BERT | Movie Reviews · TextFooler attack | attack success rate: 98.89 -> 57.69 | Improved |
| Use AT-BMC with BERT-base instead of vanilla BERT | Movie Reviews · TextBugger attack | attack success rate: 90.00 -> 26.92 | Improved |
| Use AT-BMC with BERT-base instead of vanilla BERT | Movie Reviews · PWWS attack | attack success rate: 96.67 -> 26.92 | Improved |
| Add label embedding to condition rationale extraction on the predicted label (ablation) | AT-BMC · Movie Reviews and MultiRC | Token F1: 49.8±0.51 -> 50.6±0.54; 56.2±0.47 -> 57.3±0.34 | Improved |
| Add discrete adversarial examples outside rationale spans (ablation) | AT-BMC · Movie Reviews | Token F1: 48.0±0.34 -> 50.6±0.54 | Improved |
| Add span match loss (ablation) | AT-BMC · Movie Reviews | Token F1: 47.4±0.45 -> 50.6±0.54 | Improved |
| Add virtual adversarial training in embedding space (ablation) | AT-BMC · MultiRC | accuracy: 63.6±0.47 -> 67.7±0.49 | Improved |
| Use mixed adversarial training rather than removing it (ablation) | AT-BMC · MultiRC | accuracy: 63.9±0.41 -> 67.7±0.49 | Improved |
Closest papers by meaning
- Evaluations and Methods for Explanation through Robustness Analysis (ICLR 2021 · similarity 0.88)
- Adversarial Training for Improving Model Robustness? Look at Both Prediction and Interpretation (AAAI 2022 · similarity 0.87)
- Fooling Explanations in Text Classifiers (ICLR 2022 · similarity 0.87)
- Learning Robust Rationales for Model Explainability: A Guidance-Based Approach (AAAI 2024 · similarity 0.85)
- Promoting Counterfactual Robustness through Diversity (AAAI 2024 · similarity 0.84)
- Confidence Elicitation: A New Attack Vector for Large Language Models (ICLR 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
