Continual debiasing: A bias mitigation framework for natural language

초록

Natural language understanding (NLU) systems often exploit biases in datasets for prediction, leading them to perform worse on out-of-distribution (OOD) and adversarial data. To mitigate the issue, existing debiasing methods typically down-weight the importance of losses of biased examples in the training objective during finetuning pre-trained language models (PLMs). However, they cause forgetting of learned knowledge during pre-training and hinder the model from learning intended task features in biased examples, resulting in worse in-distribution (ID) performance. In this work, we introduce a novel debiasing framework called continual debiasing to improve the trade-off between ID and OOD performance of existing debiasing methods. Specifically, the proposed framework leverages prompt tuning to PLMs' adaptation instead of full finetuning to preserve pre-trained knowledge. In addition, we perform continual debiasing with a subset of prompts that minimally affect ID performance. By doing so, we mitigate PLMs' reliance on biased features while preserving ID performance. Experimental results show that our model outperforms baselines with an average increase of 0.9 points in ID performance and 2.7 points in OOD performance. Additionally, our framework utilizes less than 0.1% of task-specific parameters compared to full finetuning approaches, leading to a substantial reduction in memory usage.

키워드