Research map / Adversarial robustness and security / Attacks on text and graphs
Adversarial Word Dilution as Text Data Augmentation in Low-Resource Regime
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 adversarial word dilution with a strict dilution constraint | BERT text classification · SST-2, k=10/20/50 | accuracy: 65.4(6.8); 72.9(5.8); 82.7(3.0) vs strongest baseline 64.4(5.8); 72.6(6.6); 82.5(2.8) | Helped |
| Use adversarial word dilution with a loose dilution constraint | BERT text classification · TREC, k=10/20/50 | accuracy: 77.1(7.1); 83.4(4.2); 90.0(3.1) vs strongest baseline 75.3(6.4); 83.3(4.6); 88.5(2.3) | Helped |
| Use adversarial word dilution | BERT text classification · SNIPS, k=10/20/50 | accuracy: 91.2(1.8); 93.8(1.0); 95.3(0.7) vs strongest baseline 91.4(1.1); 93.9(1.2); 94.8(0.7) | Mixed |
| Use the loose rather than strict dilution constraint (ablation) | AWD · SST-2, TREC, SNIPS | accuracy: 65.4(6.8) vs 65.2(7.8); 72.9(5.8) vs 73.9(5.4); 76.7(7.3) vs 77.1(7.1) | Mixed |
| Extend a pretrained dilution network to generate augmented examples without further training | AWD(strict) and AWD(loose) · SST-2, TREC, SNIPS; k=10/20/50 | accuracy: BERT+strict: 63.2(7.8), 72.7(6.2), 81.0(3.6); 75.2(9.3), 83.5(4.5), 89.2(3.2); 90.9(2.3), 93.3(1.1), 94.6(0.9) | Helped |
| Increase the strict-constraint dilution scope from too-small to moderate values (ablation) | AWD(strict) · SST-2, k=10 and k=20 | accuracy | Helped |
| Increase the strict-constraint dilution scope excessively (ablation) | AWD(strict) · SST-2, k=10 and k=20 | accuracy | Hurt |
| Use AWD-generated hard positive examples during later training | BERT text classification · SST-2, k=10 and k=20 | accuracy | Helped |
Closest papers by meaning
- Improved Text Classification via Contrastive Adversarial Training (AAAI 2022 · similarity 0.88)
- Adversarial Training Methods for Semi-Supervised Text Classification (ICLR 2017 · similarity 0.87)
- TextHoaxer: Budgeted Hard-Label Adversarial Attacks on Text (AAAI 2022 · similarity 0.87)
- UniT: A Unified Look at Certified Robust Training against Text Adversarial Perturbation (NeurIPS 2023 · similarity 0.86)
- Not All LLM-Generated Data Are Equal: Rethinking Data Weighting in Text Classification (ICLR 2025 · similarity 0.84)
- Compositional Generalization for Multi-Label Text Classification: A Data-Augmentation Approach (AAAI 2024 · 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.