Feature-Level Debiased Natural Language Understanding

Natural language understanding (NLU) models often rely on dataset biases rather than intended task-relevant features to achieve high performance specific datasets. As a result, these perform poorly datasets outside the training distribution. Some recent studies address this issue by reducing weights of biased samples during process. However, methods still encode latent in representations and neglect dynamic nature bias, which hinders model prediction. We propose an NLU debiasing method, named contrastive learning (DCT), simultaneously alleviate above problems based learning. devise debiasing, positive sampling strategy mitigate selecting least similar samples. also negative capture influence employing bias-only dynamically select most conduct experiments three benchmark Experimental results show that DCT outperforms state-of-the-art baselines out-of-distribution while maintaining in-distribution performance. verify can reduce from model's representation.