Learning from others' mistakes: Avoiding dataset biases without modeling them
arXiv:2012.01300
Abstract
State-of-the-art natural language processing (NLP) models often learn to model dataset biases and surface form correlations instead of features that target the intended underlying task. Previous work has demonstrated effective methods to circumvent these issues when knowledge of the bias is available. We consider cases where the bias issues may not be explicitly identified, and show a method for training models that learn to ignore these problematic correlations. Our approach relies on the observation that models with limited capacity primarily learn to exploit biases in the dataset. We can leverage the errors of such limited capacity models to train a more robust model in a product of experts, thus bypassing the need to hand-craft a biased model. We show the effectiveness of this method to retain improvements in out-of-distribution settings even if no particular bias is targeted by the biased model.
15 pages, 6 figures, 6 tables
References in corpus (13)
- Distilling the Knowledge in a Neural Network
- HuggingFace's Transformers: State-of-the-art Natural Language Processing
- Theoretically Principled Trade-off between Robustness and Accuracy
- SQuAD: 100,000+ Questions for Machine Comprehension of Text
- A large annotated corpus for learning natural language inference
- Well-Read Students Learn Better: On the Importance of Pre-training Compact Models
- Robustness May Be at Odds with Accuracy
- RUBi: Reducing Unimodal Biases in Visual Question Answering
- Learning and Evaluating General Linguistic Intelligence
- Gender Bias in Coreference Resolution: Evaluation and Debiasing Methods
- Are All Training Examples Created Equal? An Empirical Study
- HellaSwag: Can a Machine Really Finish Your Sentence?
- Don't Take the Easy Way Out: Ensemble Based Methods for Avoiding Known Dataset Biases