Researchers have introduced destroR, a new pipeline designed to evaluate and enhance the adversarial robustness of Bangla language transfer models. The system includes three novel meaning-preserving attack methods: a paraphrase attack, a back-translation attack, and a one-hot word-swap attack. These attacks aim to perturb inputs while maintaining semantic fidelity and fluency, thereby testing model resilience. The accompanying benchmark evaluates five transfer models, including BanglaBERT and XLM-RoBERTa, against these and other baseline attacks, revealing that word-substitution attacks are more potent than the semantically constrained ones. Adversarial training was found to improve robustness across all tested models, with the multilingual MuRIL backbone showing greater resilience than Bangla-specific models. AI
IMPACT Introduces novel methods for evaluating and improving the robustness of NLP models against adversarial attacks, particularly for under-resourced languages.
RANK_REASON The item is a research paper detailing a new benchmark and defense mechanism for NLP models. [lever_c_demoted from research: ic=1 ai=1.0]
- BAE Systems
- Bangla
- BanglaBERT
- BanglishBERT
- destroR
- IndicBERTv2
- Saadat Rafid Ahmed
- TextFooler
- XLM-RoBERTa
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
