Improving the Robustness of NLP Models

The majority of NLP datasets contain spurious patterns that are associated with the target label and are easy to learn. Models tend to focus on learning these dataset-specific spurious patterns instead of learning more generalizable patterns to solve the underlying task. As a result, while models that are trained on such datasets achieve high performances on the same data distribution, they fail on out-of-domain data distributions. In this regard, we explore innovative approaches to improve the robustness of NLP models across various datasets, tasks, and data distributions.