The Effect of Natural Distribution Shift on Question Answering Models

Abstract

We build four new test sets for the Stanford Question Answering Dataset (SQuAD) and evaluate the ability of question-answering systems to generalize to new data. Our first test set is from the original Wikipedia domain and measures the extent to which existing systems overfit the original test set. Despite several years of heavy test set re-use, we find no evidence of adaptive overfitting. The remaining three test sets are constructed from New York Times articles, Reddit posts, and Amazon product reviews and measure robustness to natural distribution shifts. Across a broad range of models, we observe average performance drops of 3.8, 14.0, and 17.4 F1 points, respectively. In contrast, a strong human baseline matches or exceeds the performance of SQuAD models on the original domain and exhibits little to no drop in new domains. Taken together, our results confirm the surprising resilience of the holdout method and emphasize the need to move towards evaluation metrics that incorporate robustness to natural distribution shifts.

References in corpus (19)

• XLNet: Generalized Autoregressive Pretraining for Language Understanding
• Machine Comprehension Using Match-LSTM and Answer Pointer
• SearchQA: A New Q&A Dataset Augmented with Context from a Search Engine
• ReasoNet: Learning to Stop Reading in Machine Comprehension
• Learning the Difference that Makes a Difference with Counterfactually-Augmented Data
• ReCoRD: Bridging the Gap between Human and Machine Commonsense Reading Comprehension
• Learning and Evaluating General Linguistic Intelligence
• Learning Recurrent Span Representations for Extractive Question Answering
• DCN+: Mixed Objective and Deep Residual Coattention for Question Answering
• A Mutual Information Maximization Perspective of Language Representation Learning
• MEMEN: Multi-layer Embedding with Memory Networks for Machine Comprehension
• Learning to Compute Word Embeddings On the Fly
• The Ladder: A Reliable Leaderboard for Machine Learning Competitions
• End-to-End Answer Chunk Extraction and Ranking for Reading Comprehension
• What do Models Learn from Question Answering Datasets?
• Exploring Question Understanding and Adaptation in Neural-Network-Based Question Answering
• Phase Conductor on Multi-layered Attentions for Machine Comprehension
• Smarnet: Teaching Machines to Read and Comprehend Like Human
• The advantages of multiple classes for reducing overfitting from test set reuse

Cited by in corpus (10)

• Learning Transferable Visual Models From Natural Language Supervision
• Knowledge Enhanced Pretrained Language Models: A Compreshensive Survey
• The Evolution of Out-of-Distribution Robustness Throughout Fine-Tuning
• E-BERT: A Phrase and Product Knowledge Enhanced Language Model for E-commerce
• Utility is in the Eye of the User: A Critique of NLP Leaderboards
• Do Offline Metrics Predict Online Performance in Recommender Systems?
• TextFlint: Unified Multilingual Robustness Evaluation Toolkit for Natural Language Processing
• MLDemon: Deployment Monitoring for Machine Learning Systems
• Understanding and Testing Generalization of Deep Networks on Out-of-Distribution Data
• QUACKIE: A NLP Classification Task With Ground Truth Explanations