LitCovid ensemble learning for COVID-19 multi-label classification

Area of Science:

• Biomedical Informatics
• Natural Language Processing
• Computational Biology

Background:

• The COVID-19 pandemic has led to an exponential increase in scientific publications.
• Manual topic extraction from this literature is inefficient and unsustainable.
• Automated methods are crucial for managing and analyzing the vast COVID-19 research output.

Purpose of the Study:

• To develop an automated method for extracting semantic topics from COVID-19 literature.
• To address the challenges of multi-label classification and imbalanced data in biomedical text mining.
• To achieve state-of-the-art performance on the BioCreative VII LitCovid Track task.

Main Methods:

• Proposed the LitCovid Ensemble Learning (LCEL) method, integrating seven transformer-based pretrained models.
• Utilized diverse biomedical knowledge and data augmentation to enhance model representation and learning.
• Implemented a novel asymmetric loss function to handle imbalanced label distributions.
• Employed an ensemble bagging strategy for final prediction generation.

Main Results:

• The LCEL method achieved state-of-the-art performance on the LitCovid dataset.
• Ensemble learning effectively combined multiple models for improved topic extraction accuracy.
• The asymmetric loss function improved model focus on positive samples in imbalanced datasets.

Conclusions:

• The LCEL method provides an effective and efficient solution for automated topic extraction in COVID-19 literature.
• Ensemble learning and specialized loss functions are valuable for biomedical text classification tasks.
• This approach aids in navigating and understanding the rapidly expanding body of COVID-19 research.