Mask-tuning: Towards Improving Pre-trained Language Models' Generalization

Pre-trained language models have the known generalization problem. This issue emerges from the pre-trained language models' learning process that heavily relies on spurious correlations, which work for the majority of training examples but do not hold in general. As a consequence, the models' performance drops substantially on out-of-distribution datasets. Previous studies proposed various solutions, including data augmentation and learning process improvement. In this paper, we present Mask-tuning, an approach that alleviates the impact of spurious correlations on the fine-tuning learning process. To achieve this goal, Mask-tuning integrates masked language training into the fine-tuning learning process. In this case, Mask-tuning perturbs the linguistic relation of downstream tasks' training examples and computes masked language training loss. Then, the perturbed examples are fed into fine-tuning process to be classified based on their ground-truth label and compute the fine-tuning training loss. Afterward, Mask-tuning loss-- a weighted aggregation of masked language model training loss and fine-tuning loss-- updates the masked language model and fine-tuning through training iterations. Extensive experiments show that Mask-tuning consistently improves the pre-trained language models' generalization on out-of-distribution datasets and enhances their performance on in-distribution datasets. The source code and pre-trained models will be available on the author's GitHub page.