Multi-Level Representation Learning for Chinese Medical Entity Recognition: Model Development and Validation.

Area of Science:

• Natural Language Processing
• Artificial Intelligence in Medicine
• Biomedical Informatics

Background:

• Medical entity recognition is crucial for smart medicine development.
• Chinese medical entity recognition lags behind English due to language complexity and limited annotated corpora.
• Existing simple neural networks struggle to extract deep semantic representations from electronic medical records (EMRs).

Purpose of the Study:

• To enhance Chinese medical entity recognition performance by learning multi-level representations.
• To address the limitations of existing models in processing complex Chinese EMRs.
• To develop a robust model for scarce Chinese medical corpora.

Main Methods:

• Proposed a multi-level representation learning model for Chinese EMRs.
• Utilized Bidirectional Encoder Representation from Transformers (BERT) for initial semantic representation extraction.
• Employed a multi-head attention mechanism to capture deeper semantic information from intermediate BERT layers.
• Integrated multi-level representations for final token embedding and entity tag prediction using softmax.

Main Results:

• Achieved an F1 score of 82.11% on the newly developed Chinese EMR (CEMR) dataset.
• Further improved performance to an F1 score of 83.18% on the CCKS 2018 benchmark dataset.
• Demonstrated superior performance compared to previous methods, establishing a new state-of-the-art.

Conclusions:

• The proposed multi-level representation learning model effectively performs entity recognition in Chinese EMRs.
• The multi-head attention mechanism is valuable for extracting multi-level representations within language models for medical tasks.
• The study highlights the potential of advanced deep learning techniques for Chinese biomedical natural language processing.