An application of convolutional neural networks with salient features for relation classification.

Area of Science:

• Biomedical informatics
• Computational biology
• Natural Language Processing

Background:

• Deep learning advancements drive interest in biomedical feature extraction and classification.
• Biomedical relation classification is crucial for understanding complex biological data.
• Existing methods require optimized feature sets and deep learning algorithms.

Purpose of the Study:

• To identify the optimal combination of feature sets and hyperparameters for deep learning-based relation classification in the biomedical domain.
• To evaluate a Convolutional Neural Networks (CNN) model against established learning algorithms for this task.

Main Methods:

• Utilized PKDE4J, an entity and relation extraction tool, to obtain biomedical features (entities, relations).
• Developed and implemented a CNN-based classification model.
• Compared the CNN model's performance with widely used supervised learning algorithms.

Main Results:

• The CNN model demonstrated superior performance compared to traditional supervised algorithms.
• Achieved a weighted macro-average F1-score of 94.79% for binary classification using optimized feature combinations.
• Attained a weighted macro-average F1-score of approximately 86.95% for multi-class classification.

Conclusions:

• The proposed CNN model significantly enhances classification performance by integrating raw text with multiple, highlighted biomedical features.
• Hyperparameter tuning and optimization approaches were developed to achieve optimal model performance.
• The findings highlight the effectiveness of combining diverse feature types for improved biomedical relation classification.