Using Adversarial Attacks to Reveal the Statistical Bias in Machine Reading Comprehension Models
This work addresses the reliability of MRC models for AI researchers and practitioners, revealing critical vulnerabilities that could impact real-world applications.
The authors tackled the problem of whether machine reading comprehension models truly understand language or exploit statistical biases, by demonstrating that adversarial attacks can reduce model performance from human-level to chance-level on the RACE dataset.
Pre-trained language models have achieved human-level performance on many Machine Reading Comprehension (MRC) tasks, but it remains unclear whether these models truly understand language or answer questions by exploiting statistical biases in datasets. Here, we demonstrate a simple yet effective method to attack MRC models and reveal the statistical biases in these models. We apply the method to the RACE dataset, for which the answer to each MRC question is selected from 4 options. It is found that several pre-trained language models, including BERT, ALBERT, and RoBERTa, show consistent preference to some options, even when these options are irrelevant to the question. When interfered by these irrelevant options, the performance of MRC models can be reduced from human-level performance to the chance-level performance. Human readers, however, are not clearly affected by these irrelevant options. Finally, we propose an augmented training method that can greatly reduce models' statistical biases.