AD-Drop: Attribution Driven Dropout for Robust Language Model Finetuning

Abstract

Fine-tuning large pre-trained language models on downstream tasks is apt to suffer from overfitting when limited training data is available. While dropout proves to be an effective antidote by randomly dropping a proportion of units, existing research has not examined its effect on the self-attention mechanism. In this paper, we investigate this problem through self-attention attribution and find that dropping attention positions with low attribution scores can accelerate training and increase the risk of overfitting. Motivated by this observation, we propose Attribution-Driven Dropout (\textsc{AD-Drop}), which randomly discards some high-attribution positions to encourage the model to make predictions by relying more on low-attribution positions to reduce overfitting. We also develop a cross-tuning strategy to alternate fine-tuning and \textsc{AD-Drop} to avoid dropping high-attribution positions excessively. Extensive experiments on various benchmarks show that \textsc{AD-Drop} yields consistent improvements over baselines.~Analysis further confirms that \textsc{AD-Drop} serves as a strategic regularizer to prevent overfitting during fine-tuning.