Enriching contextualized language model from knowledge graph for biomedical information extraction.

Area of Science:

• Biomedical Informatics
• Natural Language Processing
• Artificial Intelligence

Background:

• Biomedical information extraction (BioIE) aims to structure data from biomedical texts.
• Pre-trained language models have advanced BioIE but lack external knowledge integration.
• Existing models often overlook crucial factual information from knowledge graphs.

Purpose of the Study:

• To evaluate current BioIE methods, including neural networks and language models.
• To propose an enriched contextualized language model, BioKGLM, by integrating biomedical knowledge graphs.
• To enhance the understanding and reasoning capabilities for BioIE tasks.

Main Methods:

• Evaluated vanilla neural networks, general language models, and pre-trained contextualized language models.
• Developed BioKGLM by integrating a large-scale biomedical knowledge graph.
• Employed a three-stage training procedure and various fusion strategies for knowledge injection.

Main Results:

• BioKGLM consistently outperformed state-of-the-art extraction models across multiple BioIE tasks.
• Demonstrated superior performance in named entity recognition, relation extraction, and event extraction.
• Showcased BioKGLM's ability to capture underlying relations between biomedical concepts.

Conclusions:

• Integrating biomedical knowledge graphs significantly improves BioIE performance.
• BioKGLM offers a robust framework for enhancing BioIE through structured knowledge.
• The model's ability to understand concept relations is vital for accurate biomedical information extraction.