Research map / Label-efficient and robust learning / Recognition and long-tail
Explaining the Efficacy of Counterfactually Augmented Data
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 |
|---|---|---|---|
| Train on counterfactually augmented data rather than original data alone | Sentiment classifiers · Amazon Reviews | accuracy: 79.3 vs 74.5; 78.6 vs 74.3; 71.4 vs 68.9; 83.3 vs 80.0 | Helped |
| Train on counterfactually augmented data rather than original data alone | Sentiment classifiers · Semeval 2017 (Twitter) | accuracy: 66.8 vs 63.1; 72.4 vs 63.7; 58.2 vs 50.7; 82.8 vs 72.6 | Helped |
| Train on counterfactually augmented data rather than original data alone | Sentiment classifiers · Yelp Reviews | accuracy: 85.6 vs 81.9; 86.3 vs 82.3; 73.7 vs 72.0; 86.6 vs 84.3 | Helped |
| Add noise to human-marked rationale tokens (ablation) | Sentiment classifiers trained on original reviews · Yelp | accuracy: 79.9 -> 51.5 | Hurt |
| Add noise to unmarked tokens (ablation) | SVM trained on original IMDb reviews · Yelp | accuracy: 79.9 -> 81.4 | Helped |
| Add noise to human-marked rationale tokens (ablation) | BERT fine-tuned on SNLI · RP | accuracy: 72.7 -> 51.4 | Hurt |
| Add noise to unmarked tokens (ablation) | BERT fine-tuned on SNLI · RP | accuracy: 72.7 -> 70.4 | Hurt |
| Use human-edited CAD instead of style-transfer-generated sentiment flips | Sentiment classifiers · Amazon and Yelp | accuracy: 5--16.4%; 3.7--17.8%; 4.9--21.5%; 1.9--9.5% | Helped |
| Use attention-derived spans rather than human-provided rationales to define noisy tokens (ablation) | Sentiment classifiers · IMDb and out-of-domain sentiment datasets | accuracy | Mixed |
Closest papers by meaning
- Towards Characterizing Domain Counterfactuals for Invertible Latent Causal Models (ICLR 2024 · similarity 0.89)
- Iterative Counterfactual Data Augmentation (AAAI 2025 · similarity 0.86)
- Gumbel Counterfactual Generation From Language Models (ICLR 2025 · similarity 0.86)
- Counterfactual Generative Modeling with Variational Causal Inference (ICLR 2025 · similarity 0.86)
- LeapFactual: Reliable Visual Counterfactual Explanation Using Conditional Flow Matching (NeurIPS 2025 · similarity 0.85)
- Counterfactual reasoning: an analysis of in-context emergence (NeurIPS 2025 · similarity 0.85)
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.