Towards Interpreting and Mitigating Shortcut Learning Behavior of NLU models

NAACL 2021

Published June 11, 2021

Mengnan Du, Varun Manjunatha, Rajiv Jain, Ruchi Deshpande, Franck Dernoncourt, Jiuxiang Gu, Tong Sun, Xia Hu

Recent studies indicate that NLU models are prone to rely on shortcut features for prediction. As a result, these models could potentially fail to generalize to real-world out-of-distribution scenarios. In this work, we show that the shortcut learning behavior can be explained by the long-tailed phenomenon. There are two findings : 1) Trained NLU models have strong preference for features located at the head of the long-tailed distribution, and 2) Shortcut features are picked up during very early few iterations of the model training. These two observations are further employed to formulate a measurement which can quantify the shortcut degree of each training sample. Based on this shortcut measurement, we propose a shortcut mitigation framework, to suppress the model from making over-confident predictions for samples with large shortcut degree. Experimental results on three NLU benchmarks demonstrate that our long- tailed distribution explanation accurately reflects the shortcut learning behavior of NLU models. Experimental analysis further indicates that our method can improve the generalization accuracy on OOD data, while preserving the accuracy on in distribution test data.