Are Multilingual Language Models Fragile? IBM Adversarial Attack Strategies Cut MBERT QA Performance by 85% | Synced

MBERT is more susceptible to attacks compared to BERT.
MBERT gives priority to finding the answer in certain languages, causing successful attacks even when the adversarial statement is in a different language than the question and context.
MBERT gives priority to the language of the question over the language of the context.
Augmenting the system with machine-translated data helps build a more robust system.
— Lees op syncedreview.com/2021/04/22/deepmind-podracer-tpu-based-rl-frameworks-deliver-exceptional-performance-at-low-cost-3/