Research map / Label-efficient and robust learning / Recognition and long-tail
Simpson’s Bias in NLP Training
Fei Yuan, Longtu Zhang, Huang Bojun, Yaobo Liang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 BERT with Dice loss instead of CE loss | BERT-base-uncased · MRPC dev | F1: 89.78 -> 89.00 | Hurt |
| Train BERT with Dice loss instead of CE loss | BERT-base-uncased · QQP dev | F1: 87.84 -> 87.39 | Hurt |
| Train BERT-base multilingual cased with Dice loss instead of CE loss | BERT-base multilingual cased · GermEval 2014 test | F1: 86.14 -> 82.61 | Hurt |
| Train with the averaged sample-level objective rather than the corpus-level metric | Dice-loss training · MRPC | F1: + 0.78 | Hurt |
| Train with the averaged sample-level objective rather than the corpus-level metric | Dice-loss training · QQP | F1: + 0.45 | Hurt |
| Train with the averaged sample-level objective rather than the corpus-level metric | Dice-loss training · NER | F1: 3.53 point | Hurt |
| Use the alternative token-independent Dice loss formulation (ablation) | NER Dice-loss training · GermEval 2014 | Simpson's bias | Hurt |
| Use smoothing constant gamma=1 (ablation) | Dice-loss training · PSM and NER | Simpson's bias | No clear effect |
| Increase Simpson's bias significance | Dice-loss training · MRPC train | F1 | Hurt |
| Change learning rate, batch size, or model type (ablation) | Dice-loss training on MRPC · MRPC | Simpson's bias | No clear effect |
Closest papers by meaning
- An Empirical Study on Noisy Data and LLM Pretraining Loss Divergence (NeurIPS 2026 · similarity 0.86)
- Signal and Noise: A Framework for Reducing Uncertainty in Language Model Evaluation (NeurIPS 2025 · similarity 0.86)
- Fine-tuning can Help Detect Pretraining Data from Large Language Models (ICLR 2025 · similarity 0.85)
- Log-Likelihood, Simpson’s Paradox, and the Detection of Machine-Generated Text (NeurIPS 2026 · similarity 0.85)
- A Hitchhiker's Guide to Scaling Law Estimation (ICML 2025 · similarity 0.85)
- On the Entropy Calibration of Language Models (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.
