Broad-coverage biomedical relation extraction with SemRep.

Area of Science:

• Biomedical Informatics
• Natural Language Processing
• Computational Linguistics

Background:

• The increasing volume of biomedical literature necessitates advanced information management tools.
• Natural Language Processing (NLP) is crucial for extracting meaningful insights from vast amounts of text.
• SemRep is an NLP system designed to identify semantic relationships within biomedical abstracts.

Purpose of the Study:

• To describe the SemRep system for semantic relation extraction.
• To evaluate SemRep's performance on manually annotated and benchmark datasets.
• To identify and analyze sources of error in the relation extraction process.

Main Methods:

• Utilized linguistic principles and UMLS domain knowledge for relation extraction.
• Evaluated SemRep on a manually annotated dataset with comprehensive error analysis.
• Assessed SemRep performance on the CDR dataset for chemical-disease relationships.

Main Results:

• Relaxed evaluation on a manual dataset yielded 0.69 precision, 0.42 recall, and 0.52 F1 score.
• Named entity recognition/normalization was the primary error source (26.9%).
• On the CDR corpus, SemRep achieved 0.90 precision, 0.24 recall, and 0.38 F1 score, improving to 0.50 F1 for sentence-bound relationships.

Conclusions:

• SemRep serves as a robust baseline system for extracting biomedical semantic relations.
• It forms the foundation for SemMedDB, a large-scale biomedical knowledge graph.
• SemRep has demonstrated impact in clinical decision making, drug repurposing, and hypothesis generation.