Pay Attention to Real World Perturbations! Natural Robustness Evaluation in Machine Reading Comprehension

TL;DR

This paper introduces a framework for evaluating the robustness of machine reading comprehension models against real-world textual perturbations using Wikipedia edit history, revealing significant performance drops and potential mitigation strategies.

Contribution

It presents a novel natural perturbation evaluation method for MRC models, highlighting the gap between synthetic and real-world robustness and exploring training-based improvements.

Findings

Natural perturbations cause performance degradation in state-of-the-art models.

Models like Flan-T5 and LLMs inherit errors from natural perturbations.

Training on perturbed data improves robustness but does not fully close the gap.

Abstract

As neural language models achieve human-comparable performance on Machine Reading Comprehension (MRC) and see widespread adoption, ensuring their robustness in real-world scenarios has become increasingly important. Current robustness evaluation research, though, primarily develops synthetic perturbation methods, leaving unclear how well they reflect real life scenarios. Considering this, we present a framework to automatically examine MRC models on naturally occurring textual perturbations, by replacing paragraph in MRC benchmarks with their counterparts based on available Wikipedia edit history. Such perturbation type is natural as its design does not stem from an arteficial generative process, inherently distinct from the previously investigated synthetic approaches. In a large-scale study encompassing SQUAD datasets and various model architectures we observe that natural perturbations result in performance degradation in pre-trained encoder language models. More worryingly, these state-of-the-art Flan-T5 and Large Language Models (LLMs) inherit these errors. Further experiments demonstrate that our findings generalise to natural perturbations found in other more challenging MRC benchmarks. In an effort to mitigate these errors, we show that it is possible to improve the robustness to natural perturbations by training on naturally or synthetically perturbed examples, though a noticeable gap still remains compared to performance on unperturbed data.