Reproduce “Unifying Model Explainability and Robustness for Joint Text Classification and Rationale Extraction” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Dongfang Li, Baotian Hu, Qingcai Chen, Tujie Xu, Jingcong Tao, Yunan Zhang
Researchers at the Harbin Institute of Technology (Shenzhen) developed AT-BMC, a unified framework that jointly performs text classification and rationale extraction, while simultaneously enhancing model robustness. This system achieves improved classification accuracy and rationale F1 scores while significantly reducing the attack success rate of adversarial perturbations.
Key insights
- Unifying model explainability and adversarial robustness within a single framework leads to synergistic improvements in both dimensions for text classification models.
- Employing mixed adversarial training across both discrete text and continuous embedding spaces effectively enhances robustness against diverse attacks while preserving explanatory quality.
- Explicitly modeling rationale span boundaries through a Boundary Match Constraint significantly improves the accuracy and coherence of extracted rationales, aligning them better with human intuition.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.