Leveraging Semantic Type Dependencies for Clinical Named Entity Recognition.

Area of Science:

• Natural Language Processing
• Clinical Informatics
• Biomedical Text Mining

Background:

• Clinical relation extraction often uses semantic types from knowledge bases.
• Existing methods may not fully exploit the relationships between these semantic types.

Purpose of the Study:

• To improve clinical relation extraction by utilizing domain-specific semantic type dependencies.
• To develop and evaluate a novel matrix encoding for processing multiple dependencies in Named Entity Recognition (NER).

Main Methods:

• Encoding relations between Unified Medical Language System (UMLS) concepts and other sentence tokens.
• Implementing and comparing NER architectures (BiLSTM-CRF, BiLSTM-GCN-CRF) with various clinical embeddings (BERT, BioBERT, UMLSBert).
• Introducing a matrix encoding for multi-dependency processing in a single pass.

Main Results:

• Domain-specific semantic type dependencies can significantly improve NER effectiveness in clinical datasets.
• The proposed matrix encoding is the first to handle more than three dependencies simultaneously for NER.
• Experimental results demonstrate performance gains with the inclusion of semantic type dependencies.

Conclusions:

• Incorporating domain-specific semantic type dependencies is a valuable strategy for enhancing clinical NER.
• The novel matrix encoding offers a more efficient way to leverage complex dependency information.
• This research advances the field of biomedical text mining by providing improved methods for clinical relation extraction.