Refining Language Models with Compositional Explanations

Huihan Yao, Ying Chen, Qinyuan Ye +2

Pre-trained language models have been successful on text classification tasks, but are prone to learning spurious correlations from biased datasets, and are thus vulnerable when ma…