Research map / Adversarial robustness and security / Attacks on text and graphs
Token-Aware Virtual Adversarial Training in Natural Language Understanding
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 token-aware virtual adversarial perturbations | BERT-base · RTE test | accuracy: 70.1 -> 71.0 | Helped |
| Use token-aware virtual adversarial perturbations | BERT-base · CoNLL2003 | F1: 94.9 -> 95.4 | Helped |
| Use token-aware virtual adversarial perturbations | BERT · IMDB | accuracy: 95.4 -> 95.7 | Helped |
| Initialize token perturbations from the accumulated perturbation vocabulary rather than randomly (ablation) | TA-VAT · RTE dev | accuracy: 73.0 -> 74.0 | Helped |
| Apply token-level normalization instead of instance-level Frobenius normalization (ablation) | TA-VAT · RTE dev | accuracy: 72.5 -> 74.0 | Helped |
| Add token-aware initialization and token-level normalization (ablation) | FreeLB-ReImp · RTE dev | accuracy: 70.0 -> 74.0 | Helped |
| Initialize normal fine-tuning with a task-learned perturbation vocabulary (ablation) | BERT-base fine-tuning · RTE dev | accuracy: 63.5 -> 72.0 | Helped |
| Initialize normal fine-tuning with a perturbation vocabulary learned on another task (ablation) | BERT-base fine-tuning · RTE dev | accuracy: 63.5 -> 69.0 | Helped |
| Include special tokens in the perturbation vocabulary (ablation) | TA-VAT · RTE dev | accuracy: 73.0 -> 74.0 | Helped |
| Train with TA-VAT rather than standard BERT | BERT-base · IMDB under TextFooler attack | attacked accuracy: 13.6 -> 15.8 | Helped |
| Train with TA-VAT rather than standard BERT | BERT-base · IMDB under TextFooler attack | query number: 1134 -> 2093 | Helped |
| Train with TA-VAT rather than standard BERT | BERT-base · MNLI with 2000 training examples | performance | Helped |
Closest papers by meaning
- Adversarial Training Methods for Semi-Supervised Text Classification (ICLR 2017 · similarity 0.90)
- TextGrad: Advancing Robustness Evaluation in NLP by Gradient-Driven Optimization (ICLR 2023 · similarity 0.87)
- Toward Adversarial Training on Contextualized Language Representation (ICLR 2023 · similarity 0.87)
- Revisiting Character-level Adversarial Attacks for Language Models (ICML 2024 · similarity 0.87)
- Efficient Adversarial Training in LLMs with Continuous Attacks (NeurIPS 2024 · similarity 0.86)
- Enhancing NLU in Large Language Models Using Adversarial Noisy Instruction Tuning (AAAI 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.