Improving BERT with Self-Supervised Attention
This work addresses a common issue in NLP for practitioners using BERT on limited data, offering an incremental improvement to existing fine-tuning methods.
The paper tackles the problem of BERT overfitting on small datasets during fine-tuning by proposing Self-Supervised Attention (SSA), which iteratively generates token-level attention labels to improve generalization, resulting in significant performance gains across multiple public datasets.
One of the most popular paradigms of applying large pre-trained NLP models such as BERT is to fine-tune it on a smaller dataset. However, one challenge remains as the fine-tuned model often overfits on smaller datasets. A symptom of this phenomenon is that irrelevant or misleading words in the sentence, which are easy to understand for human beings, can substantially degrade the performance of these finetuned BERT models. In this paper, we propose a novel technique, called Self-Supervised Attention (SSA) to help facilitate this generalization challenge. Specifically, SSA automatically generates weak, token-level attention labels iteratively by probing the fine-tuned model from the previous iteration. We investigate two different ways of integrating SSA into BERT and propose a hybrid approach to combine their benefits. Empirically, through a variety of public datasets, we illustrate significant performance improvement using our SSA-enhanced BERT model.