BioRED

BioRED provides a document-level biomedical relation extraction dataset that annotates genes/proteins, diseases, and chemicals across diverse relation pairs and labels each relation for novelty to support development and evaluation of automated RE and NER systems.

Key Features:

• Multiple Entity Types: Includes annotations for genes/proteins, diseases, and chemicals.
• Diverse Relation Pairs: Contains various relation pairs, including gene-disease and chemical-chemical interactions.
• Document-Level Scope: Operates at the document level to capture context-rich relationships beyond single sentences.
• Novelty Annotation: Labels each relation as a novel finding or previously known background knowledge.

Scientific Applications:

• Relation Extraction development: Enables training and benchmarking of RE systems, including models based on Bidirectional Encoder Representations from Transformers (BERT).
• Named Entity Recognition benchmarking: Serves as a resource for evaluating NER performance on genes/proteins, diseases, and chemicals.
• Novelty analysis in biomedical text: Facilitates research into distinguishing novel findings from background knowledge in biomedical literature.

Methodology:

Constructed after a review of existing NER and RE datasets, assembled from 600 PubMed abstracts, and evaluated via benchmarking experiments with state-of-the-art methods that reported high NER performance while highlighting challenges in extracting novel relations.