NLM DIR Seminar Schedule

Abstract:

Previous studies on biomedical relation extraction (RE) typically focus on extracting binary relations between two entities from a single sentence. However, complex inter-sentence relations involving multiple entity pairs, such as drug-protein and protein-disease, are commonly seen in the biomedical literature. In this talk, I will first introduce the characteristics of sentence-level RE and use the BioCreative VII DrugProt task to showcase a general text classification framework for sentence-level RE. The second part will introduce a new document-level dataset called BioRED, which covers six concept types (cell line, chemical, disease, gene, species, and variant) and eight relation pairs (e.g., chemical-disease, chemical-gene, chemical-chemical) in 600 MEDLINE abstracts. In total, BioRED consists of 20,000 entity and 6,000 relation annotations. The BioRED dataset is currently being used for developing and evaluating state-of-the-art relation extraction methods at the LitCoin natural language processing (NLP) challenge.