Robustness of Named Entity Recognition Models

Transformer based models are gaining popularity on most of the Natural Language Processing (NLP) including sequence labeling task such like Named Entity Recognition (NER). The previous (pre-transformers) state of the art approaches for NER task where models based on Conditional Random Fields (CRF). Adversarial attacks, which involve trying to deceive the model by creating adversarial examples with subtle, almost imperceptible changes, are increasingly posing a threat to all NLP models, including those which solve sequence labeling tasks. When deciding which architecture to use for NER model, robustness of approach could be one of the deciding factors. We are going to examine named entity recognition trained models resilience, focusing on differences between CRF and Transformer based approaches.