Bag of Lies: Robustness in Continuous Pre-training BERT

TL;DR

This paper investigates the robustness of continuous pre-training of BERT, especially in the context of new entity knowledge like COVID-19, revealing surprising resilience against misinformation and introducing a new dataset.

Contribution

It provides insights into how continuous pre-training affects BERT's entity knowledge and robustness, and introduces a new dataset with AI-generated false texts.

Findings

Pre-training does not degrade performance under adversarial input manipulations.

Continuous pre-training can sometimes improve downstream task performance.

The model shows robustness against misinformation during continuous pre-training.

Abstract

This study aims to acquire more insights into the continuous pre-training phase of BERT regarding entity knowledge, using the COVID-19 pandemic as a case study. Since the pandemic emerged after the last update of BERT's pre-training data, the model has little to no entity knowledge about COVID-19. Using continuous pre-training, we control what entity knowledge is available to the model. We compare the baseline BERT model with the further pre-trained variants on the fact-checking benchmark Check-COVID. To test the robustness of continuous pre-training, we experiment with several adversarial methods to manipulate the input data, such as training on misinformation and shuffling the word order until the input becomes nonsensical. Surprisingly, our findings reveal that these methods do not degrade, and sometimes even improve, the model's downstream performance. This suggests that continuous pre-training of BERT is robust against misinformation. Furthermore, we are releasing a new dataset, consisting of original texts from academic publications in the LitCovid repository and their AI-generated false counterparts.