Counterfactuals and Dynamic Sampling Combat Spurious Correlations in NLI

This research addresses a critical challenge in Natural Language Inference (NLI) by proposing a novel method to mitigate spurious correlations. The use of LLM-synthesized counterfactuals and dynamic balanced sampling represents a promising approach to improve the robustness and generalization of NLI models.

• •Addresses the problem of spurious correlations in NLI.
• •Employs LLM-synthesized counterfactuals for data augmentation.
• •Utilizes dynamic balanced sampling for training.