Contributions
- Identify strong dataset artifacts in FEVER: n-grams (especially negations) are highly predictive of labels independent of evidence.
- Introduce a symmetric test set construction method that forces models to reason over evidence by creating claim-evidence pairs with inverse relations.
- Propose a re-weighting regularization procedure that flattens the correlation between claim n-grams and labels in the training data.
Method
The paper diagnoses dataset bias through Local Mutual Information (LMI) analysis, showing that bigrams like "did not," "refused to," and "failed to" are strong predictors of the REFUTES label. They then construct a symmetric test set: for each original REFUTES claim-evidence pair, they manually generate a synthetic pair expressing a contrary fact—the new evidence supports the original claim, and the new claim is supported by the original evidence. By construction, knowing any n-gram's presence gives 50% probability of either label.
For training, they introduce a re-weighting scheme that assigns balancing weights α(i) to training instances to minimize the bias term:
$$b^c_j = \frac{\sum^n_{i=1} I[w^{(i)}_j](1 + α^{(i)})I[y^{(i)}=c]}{\sum^n_{i=1} I[w^{(i)}_j](1 + α^{(i)})}$$
This flattens the distribution of "give-away" phrases across labels. The training objective is reweighted from $\sum_i L(x^{(i)}, y^{(i)})$ to $\sum_i (1 + α^{(i)})L(x^{(i)}, y^{(i)})$.
Results
On the original FEVER dev set, three classifiers (NSMN, ESIM, BERT) achieve 58.7%–86.2%. On the symmetric test set, all drop significantly: NSMN to 58.7%, ESIM to 55.9%, BERT to 58.3%. The re-weighting method recovers 3.3–3.4 percentage points on the symmetric set (ESIM: 55.9% → 59.3%; BERT: 58.3% → 61.6%), at a small cost on the biased dev set (showing it trades off exploiting bias for robustness).
Notes
Strengths: The symmetric test set is an elegant, practical way to force evidence-dependent reasoning. The LMI analysis clearly isolates the bias source. The re-weighting method is simple, model-agnostic, and well-motivated. The paper makes a real contribution to understanding why FEVER is vulnerable to shortcut learning.
Limitations: The symmetric test set is small (956 pairs) and created manually only for validation, not for training. The regularization method requires computing optimal weights as a pre-processing step, which may not scale. The paper doesn't explore whether similar biases exist in other popular fact-checking datasets, though it cites evidence that they do.
Follow-ups: This work motivated later efforts in dataset debiasing (e.g., adversarial data augmentation, hard-negative mining) and evaluation protocols that stress-test fact-checking models on distribution-shifted evidence.