MMBERT: a unified framework for biomedical named entity recognition.

Area of Science:

• Natural Language Processing (NLP)
• Biomedical Informatics
• Machine Learning

Background:

• Biomedical Named Entity Recognition (NER) is crucial but challenging due to limited datasets and complex entities.
• Existing methods struggle with the unique characteristics of biomedical text.
• The scarcity of annotated biomedical NER datasets hinders model development.

Purpose of the Study:

• To propose a novel framework, MMBERT, for enhanced biomedical NER.
• To address the challenges of data scarcity and entity complexity in biomedical NER.
• To improve the generalization capability and prediction accuracy of biomedical NER models.

Main Methods:

• Introduced ERNIE-Health, a Chinese language representation model pre-trained on biomedical corpora.
• Utilized BERT and CW-LSTM for joint feature vectors of word pair relations.
• Employed multi-granularity 2D convolution for refining word pair representations and relationships.
• Designed a Convolutional Neural Network (CNN) and co-predictor for improved generalization and accuracy.

Main Results:

• The proposed MMBERT framework achieved superior performance compared to several baseline models.
• Experiments were conducted on three benchmark datasets, demonstrating the model's effectiveness.
• The model successfully addressed the challenges of data scarcity and entity complexity.

Conclusions:

• MMBERT represents a significant advancement in biomedical NER.
• The integration of ERNIE-Health and advanced deep learning techniques enhances model performance.
• The framework offers a robust solution for extracting valuable information from biomedical texts.