The Effect of Natural Distribution Shift on Question Answering Models

TL;DR

This paper evaluates question-answering models' robustness to natural distribution shifts by creating new test sets from different domains, revealing significant performance drops and highlighting the need for more robust evaluation metrics.

Contribution

The paper introduces four new test sets for SQuAD from diverse domains to assess models' generalization and robustness to distribution shifts.

Findings

Models experience average F1 drops of 3.8, 14.0, and 17.4 across new domains.

Human performance remains stable across original and new domains.

The holdout method shows surprising resilience despite distribution shifts.

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.