Don't Take the Easy Way Out: Ensemble Based Methods for Avoiding Known Dataset Biases

Abstract

State-of-the-art models often make use of superficial patterns in the data that do not generalize well to out-of-domain or adversarial settings. For example, textual entailment models often learn that particular key words imply entailment, irrespective of context, and visual question answering models learn to predict prototypical answers, without considering evidence in the image. In this paper, we show that if we have prior knowledge of such biases, we can train a model to be more robust to domain shift. Our method has two stages: we (1) train a naive model that makes predictions exclusively based on dataset biases, and (2) train a robust model as part of an ensemble with the naive one in order to encourage it to focus on other patterns in the data that are more likely to generalize. Experiments on five datasets with out-of-domain test sets show significantly improved robustness in all settings, including a 12 point gain on a changing priors visual question answering dataset and a 9 point gain on an adversarial question answering test set.

arXiv commentIn EMNLP 2019

References in corpus (11)

• Bidirectional Attention Flow for Machine Comprehension
• Enhanced LSTM for Natural Language Inference
• VQA: Visual Question Answering
• Highway Networks
• Think you have Solved Question Answering? Try ARC, the AI2 Reasoning Challenge
• Machine Comprehension Using Match-LSTM and Answer Pointer
• ReCoRD: Bridging the Gap between Human and Machine Commonsense Reading Comprehension
• Women also Snowboard: Overcoming Bias in Captioning Models
• Blindfold Baselines for Embodied QA
• Balanced Datasets Are Not Enough: Estimating and Mitigating Gender Bias in Deep Image Representations
• Unlearn Dataset Bias in Natural Language Inference by Fitting the Residual

Cited by in corpus (18)

• On the Value of Out-of-Distribution Testing: An Example of Goodhart's Law
• Learning from others' mistakes: Avoiding dataset biases without modeling them
• Unshuffling Data for Improved Generalization
• Invariant Rationalization
• Removing Bias in Multi-modal Classifiers: Regularization by Maximizing Functional Entropies
• CrossAug: A Contrastive Data Augmentation Method for Debiasing Fact Verification Models
• HybridQA: A Dataset of Multi-Hop Question Answering over Tabular and Textual Data
• Jigsaw-VAE: Towards Balancing Features in Variational Autoencoders
• Evaluating Gender Bias in Natural Language Inference
• Counterfactual Variable Control for Robust and Interpretable Question Answering
• Toward Learning Human-aligned Cross-domain Robust Models by Countering Misaligned Features
• DQI: A Guide to Benchmark Evaluation
• Our Evaluation Metric Needs an Update to Encourage Generalization
• Learning content and context with language bias for Visual Question Answering
• Focus-and-Expand: Training Guidance Through Gradual Manipulation of Input Features
• On the Significance of Question Encoder Sequence Model in the Out-of-Distribution Performance in Visual Question Answering
• X-GGM: Graph Generative Modeling for Out-of-Distribution Generalization in Visual Question Answering
• Mitigating the Position Bias of Transformer Models in Passage Re-Ranking