[Paper Review] Why Machine Reading Comprehension Models Learn Shortcuts?

Yuxuan Lai, Chen Zhang+3 more|arXiv (Cornell University)|Jun 2, 2021

Topic Modeling28 references4 citations

TL;DR

This paper investigates why machine reading comprehension (MRC) models learn shortcut solutions instead of genuine comprehension skills. Using two synthetic datasets with paired shortcut and challenging questions, the authors show that models learn shortcut questions earlier and more easily due to higher gradient optimization priority, and that high proportions of shortcut questions in training hinder the development of complex reasoning skills like paraphrasing.

ABSTRACT

Recent studies report that many machine reading comprehension (MRC) models can perform closely to or even better than humans on benchmark datasets. However, existing works indicate that many MRC models may learn shortcuts to outwit these benchmarks, but the performance is unsatisfactory in real-world applications. In this work, we attempt to explore, instead of the expected comprehension skills, why these models learn the shortcuts. Based on the observation that a large portion of questions in current datasets have shortcut solutions, we argue that larger proportion of shortcut questions in training data make models rely on shortcut tricks excessively. To investigate this hypothesis, we carefully design two synthetic datasets with annotations that indicate whether a question can be answered using shortcut solutions. We further propose two new methods to quantitatively analyze the learning difficulty regarding shortcut and challenging questions, and revealing the inherent learning mechanism behind the different performance between the two kinds of questions. A thorough empirical analysis shows that MRC models tend to learn shortcut questions earlier than challenging questions, and the high proportions of shortcut questions in training sets hinder models from exploring the sophisticated reasoning skills in the later stage of training.

Motivation & Objective

• To investigate why MRC models learn shortcut solutions instead of intended comprehension skills like paraphrasing or co-reference resolution.
• To address the lack of datasets labeled for shortcut vs. challenging questions, which limits empirical analysis of model learning behavior.
• To quantify how the proportion of shortcut questions in training data affects model learning dynamics and performance on challenging reasoning tasks.
• To develop methods for probing the learning mechanism behind model performance differences on shortcut versus complex reasoning questions.

Proposed method

• Design two synthetic MRC datasets where each (passage, question) pair has a shortcut version and a challenging version requiring paraphrasing or co-reference resolution.
• Ensure paired questions are nearly identical in style, length, and topic to isolate the effect of reasoning complexity.
• Propose two evaluation methods to quantitatively measure the learning difficulty of shortcut versus challenging question sets during training.
• Use gradient-based optimization analysis to track when models learn shortcut vs. challenging answers, identifying learning priority.
• Train BiDAF and BERT models on varying proportions of shortcut questions to observe training dynamics and performance shifts.
• Analyze model behavior via step-by-step accuracy tracking, showing when models first achieve correct answers on each question type.

Experimental results

Research questions

• RQ1 Why do MRC models learn shortcut solutions more readily than intended comprehension skills?
• RQ2 How does the proportion of shortcut questions in training data affect model learning of complex reasoning skills?
• RQ3 At what stage of training do models learn shortcut questions compared to challenging ones?
• RQ4 What role does gradient optimization play in the prioritization of shortcut learning over deeper reasoning?
• RQ5 Can we quantitatively measure the learning difficulty difference between shortcut and challenging questions?

Key findings

• Models learn shortcut questions significantly earlier than challenging questions, with BERT correctly answering a shortcut question in as few as 380 training steps.
• The challenging paraphrasing-based questions are not answered correctly until 630 steps under the same training setup, indicating a clear learning delay.
• Higher proportions of shortcut questions in training data lead to models relying excessively on shortcuts, reducing motivation to learn complex reasoning in later training stages.
• Gradient-based optimizers prioritize fitting shortcut patterns early in training, which hinders exploration of sophisticated reasoning skills.
• The synthetic datasets successfully isolate reasoning complexity, enabling controlled analysis of model learning behavior.
• The results empirically confirm that models tend to learn easier decision boundaries first, with shortcut questions representing such easier boundaries in MRC.