Enhancing Generalizability in Biomedical Entity Recognition: Self-Attention PCA-CLS Model.

Area of Science:

• Biomedical Natural Language Processing
• Computational Biology
• Data Mining

Background:

• Biomedical entity recognition is crucial for data mining.
• Traditional methods struggle with unannotated data and domain generalization.
• Deep learning models show promise but face challenges with out-of-vocabulary terms.

Purpose of the Study:

• To develop a novel deep learning model for improved biomedical entity recognition.
• To address challenges in domain generalization and out-of-vocabulary terms.
• To evaluate the model's performance across diverse biomedical datasets.

Main Methods:

• Developed the PCA-CLS (Position and Contextual Attention with CNN-LSTM-Softmax) model.
• Integrated global self-attention and character-level convolutional neural network techniques.
• Evaluated the model on eight distinct biomedical datasets (genes, drugs, diseases, species).

Main Results:

• The PCA-CLS model demonstrated superior performance compared to state-of-the-art models.
• Achieved high F-scores across multiple datasets, including 88.19% on BC2GM and 91.59% on s800.
• Effectively handled out-of-vocabulary challenges and improved domain generalization.

Conclusions:

• The PCA-CLS model offers a robust solution for biomedical entity recognition.
• Attention mechanisms and CNN-LSTM-Softmax integration are key to its success.
• The model shows significant potential for advancing biomedical text mining applications.