BiasPAD: A Bias-Progressive Auto-Debiasing Framework

Hao Huang, Tianyi Zhou, Guodong Long, Tao Shen, Jing Jiang, Chengqi Zhang

Abstract: While large pre-trained language models have made great strides in natural language understanding benchmarks, recent studies have found that models rely more on the superficial or short-cut features to make predictions. In this paper, we study how to progressively and automatically detect and filter the biased data to train a robust debiased model for NLU tasks. Rather than focusing on the human-predefined biases or biases captured by a bias-only model of limited-capacity, we introduce a new debiasing framework, called Bias-Progressive Auto-Debiasing (BiasPAD), based on two observations: i) the higher the proportion of bias in the training data, the more biased the model will be, and ii) a more biased model has higher confidence in predicting the bias. The framework progressively trains a bias-only model by using the most biased samples detected in the previous epoch, which ensures a more biased model and leads to a robust debiased model. The extensive experiments demonstrate the effectiveness of the proposed framework on several challenging NLU datasets, where on HANS, we achieve 5% accuracy improvement.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Applications (eg, speech processing, computer vision, NLP)