AI Research Proposes Product-of-Experts Training to Reduce Dataset Artifacts in Natural Language Inference

A new AI training method called Product-of-Experts (PoE) has been developed to address dataset artifacts in Natural Language Inference models. Neural NLI models frequently overfit to these artifacts rather than engaging in genuine reasoning, as demonstrated by a hypothesis-only model achieving 57.7% accuracy on the SNLI dataset. This reveals strong spurious correlations, with 38.6% of baseline errors attributable to such artifacts. The PoE approach works by downweighting examples where biased models exhibit excessive confidence. In practical application, PoE maintains near-identical accuracy at 89.10% compared to the baseline 89.30%, while simultaneously reducing bias reliance by 4.71%—shifting bias agreement from 49.85% to 45%. An ablation study identified lambda = 1.5 as the optimal parameter for balancing debiasing efforts with accuracy preservation. Despite these improvements, behavioral testing continues to uncover persistent challenges with negation and numerical reasoning within the models. The research was published on arXiv under the Computer Science > Computation and Language category with the identifier 2604.19069.