Improving BERT With Self-supervised Attention

Chen Yiren, Kou Xiaoyu, Bai Jiangang, Tong Yunhai. Arxiv 2020

[Paper]    
Attention Mechanism BERT Model Architecture Training Techniques

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.

Similar Work