Benchmarking Biomedical Nested NER and Relation Extraction Models

Abstract

The Open EPPI corpus comprises 151 full-text papers annotated by domain experts for entity mentions, protein-protein interactions (PPIs), and normalisation of entities to publicly available ontologies. The corpus is publicly available at [ANON]. We benchmark recent nested NER and relation extraction models. Results show that, although existing nested NER models achieve good performance on outermost and innermost entity mentions, they struggle with other types of nested mentions. Benchmark results for relation extraction show substantial room for improvement with precision under 70 and recall around 40 to 52.

Tasks

Reproductions