Flat and Nested Negation and Uncertainty Detection with PubMed BERT

Abstract

Negation and uncertainty detection is an oft-studied challenge in biomedical NLP. Annotation style for the task has not been standardized and as such, the existing datasets not only vary in domain but require various algorithmic designs due to their structural differences. We present a new negation detection dataset in two versions from clinical publications. We further developed two BERT-based models to evaluate on each dataset version. Both models treat the task as a token-level multi-class classification task, one of which is capable of assigning more than one label per token in the case of recursive nesting. Our models achieve F1 scores of 76% and 72% on the development and test sets, respectively.

Tasks

Reproductions