Is metadata of articles about COVID-19 enough for multilabel topic classification task?

Area of Science:

• Bibliometrics
• Computational Linguistics
• Medical Informatics

Background:

• The exponential growth of COVID-19 literature necessitates automated methods for topic classification.
• Manual curation of biomedical literature, such as LitCovid, faces scalability challenges.
• Existing multilabel classification methods struggle with label correlation and imbalance.

Purpose of the Study:

• To develop a robust multilabel topic classification framework for COVID-19 research articles.
• To evaluate the discriminative power of different article components (full text, MeSH, biological entities, metadata) for topic classification.
• To compare the performance of the proposed framework against existing methods.

Main Methods:

• Development of a novel multilabel classification framework incorporating pretrained models.
• The framework addresses label correlation and imbalance issues.
• Experiments conducted on the Enriched BC7-LitCovid and Hallmarks of Cancer corpora.

Main Results:

• The proposed framework demonstrated superior performance and robustness in multilabel topic classification.
• Article metadata (title, abstract, keyword, journal name) proved to contain significant discriminative information.
• Full texts and Medical Subject Headings (MeSH) provided marginal improvements over metadata alone.

Conclusions:

• The developed framework effectively handles the complexities of topic classification in large biomedical datasets.
• Metadata is a crucial feature for understanding the thematic content of COVID-19 research.
• While full text and MeSH offer some benefits, their impact on classification performance is limited compared to metadata.