No More Distractions: an Adaptive Up-Sampling Algorithm to Reduce Data Artifacts
Han Chen
TL;DR
Spurious token-label correlations can cause NLP models to perform well on benchmarks without truly understanding semantics. The paper analyzes SNLI data to quantify token-label bias using metrics like $p^{*}$ and $z^{*}$, identifies top biased tokens, and demonstrates that models disproportionately succeed on records containing majority-label tokens. It then introduces AUDAC, an adaptive upsampling method that iteratively balances $p(label|token)$ toward a uniform distribution across labels for the top biased tokens, without requiring human edits. In experiments, applying AUDAC yields a modest but meaningful improvement in overall accuracy (from 89.149% to 89.667%) and increases performance on corrected-token subsets, while substantially reducing artifact indicators, highlighting a practical approach to mitigate data artifacts and improve generalization in NLP.
Abstract
Researchers recently found out that sometimes language models achieve high accuracy on benchmark data set, but they can not generalize very well with even little changes to the original data set. This is sometimes due to data artifacts, model is learning the spurious correlation between tokens and labels, instead of the semantics and logic. In this work, we analyzed SNLI data and visualized such spurious correlations. We proposed an adaptive up-sampling algorithm to correct the data artifacts, which is simple and effective, and does not need human edits or annotation. We did an experiment applying the algorithm to fix the data artifacts in SNLI data and the model trained with corrected data performed significantly better than the model trained with raw SNLI data, overall, as well as on the subset we corrected.