Abstract
Question answering (QA) models often rely on "shortcuts"—spurious correlations—rather than true comprehension. This paper investigates whether models exhibit preferences when multiple shortcuts are available. Through behavioral testing, loss landscape visualization, and information-theoretic analysis, the authors demonstrate that models favor shortcuts with flatter, deeper loss landscapes and lower Minimum Description Length (MDL). These findings provide a quantitative framework for understanding and mitigating dataset bias.
Introduction
Deep learning models are often "lazy learners," exploiting shortcuts like answer position or word matching instead of mastering complex reading comprehension. While the existence of these shortcuts is well-documented, real-world datasets typically contain multiple competing shortcuts. This research addresses the hierarchy of "learnability": which shortcuts do models prefer to learn, and why? Understanding this is critical for developing effective debiasing strategies, as current methods often treat all biases as equal, which might leave the most "tempting" shortcuts untouched.
Methodology
The researchers employed three primary strategies to analyze BERT and RoBERTa models:
- Behavioral Tests: Specialized "biased training sets" were constructed where every example was solvable by three distinct shortcuts simultaneously. Models were then evaluated on "unbiased" test sets where only one specific shortcut was applicable to determine which one the model actually prioritized.
- Loss Landscape Visualization: The loss surface in the model's parameter space was mapped to identify geometric features, such as flatness and depth, associated with preferred shortcuts.
- Rissanen Shortcut Analysis (RSA): Using the concept of Minimum Description Length (MDL), the authors quantified the "effort" (in bits) required to learn a shortcut. Lower MDL indicates a shortcut is easier for the model to internalize.
Key Insights
- The Hierarchy of Laziness: In extractive QA, models overwhelmingly prefer the Answer-Position shortcut (e.g., looking at the first sentence) over word-matching or entity-type matching. In multiple-choice QA, Word-label Correlations eventually become the preferred shortcut over simple lexical overlap.
- Geometric Preference: Preferred shortcuts inhabit flatter and deeper regions of the loss landscape. In deep learning, flatter minima are generally easier for optimizers to locate and maintain, explaining the model's gravitation toward them.
- Information-Theoretic Predictability: There is a strong correlation between a shortcut’s MDL and its preference. Shortcuts that allow for a "shorter description" of the data are learned first.
- Non-Uniform Mitigation: The amount of "anti-shortcut" data (examples where the shortcut fails) required to debias a model depends on the shortcut's learnability. Highly learnable shortcuts require significantly more counter-examples to "unlearn."
Results
On SQuAD 1.1, the Answer-Position shortcut remained dominant throughout the entire training process, with F1 scores significantly higher than those for word or type matching. Quantitative analysis via MDL confirmed this: the Position shortcut scored 4.65 kbits for BERT, compared to 5.75 kbits for Type matching. This gap demonstrates that the Position shortcut is computationally "cheaper" to learn. Furthermore, the study found that simply adding a small number of hard examples is insufficient for highly learnable shortcuts; they require a much higher proportion of clean data to achieve balanced performance.
Limitations and Future Work
The study is limited to shortcuts already identified in existing literature; latent shortcuts may exist that are even more preferred. Additionally, while BERT and RoBERTa were tested, newer architectures like T5 or GPT-4 might exhibit different inductive biases. Future research could leverage the MDL-based learnability metric to automatically identify prone areas in datasets, streamlining the creation of robust training sets without manual trial-and-error.
Summary updated