An Empirical Study on Robustness to Spurious Correlations using Pre-trained Language Models
This work addresses robustness issues in NLP models for researchers and practitioners, but it is incremental as it builds on existing findings about spurious correlations and MTL.
The study investigated how pre-trained language models handle spurious correlations in datasets, finding they rely on minority counterexamples for robustness and perform poorly when these are scarce. The researchers proposed multi-task learning (MTL) with appropriate auxiliary tasks, which improved performance on challenging examples by up to 15% without harming in-distribution accuracy in natural language inference and paraphrase identification tasks.
Recent work has shown that pre-trained language models such as BERT improve robustness to spurious correlations in the dataset. Intrigued by these results, we find that the key to their success is generalization from a small amount of counterexamples where the spurious correlations do not hold. When such minority examples are scarce, pre-trained models perform as poorly as models trained from scratch. In the case of extreme minority, we propose to use multi-task learning (MTL) to improve generalization. Our experiments on natural language inference and paraphrase identification show that MTL with the right auxiliary tasks significantly improves performance on challenging examples without hurting the in-distribution performance. Further, we show that the gain from MTL mainly comes from improved generalization from the minority examples. Our results highlight the importance of data diversity for overcoming spurious correlations.