A BERT-based ensemble learning approach for the BioCreative VII challenges: full-text chemical identification and...

Area of Science:

• Biomedical Natural Language Processing
• Machine Learning in Bioinformatics
• Computational Chemistry

Background:

• Biomedical text mining requires advanced models for accurate information extraction.
• Existing Bidirectional Encoder Representations from Transformers (BERT) models offer potential but require domain-specific optimization.
• The BioCreative VII Challenge presents specific tracks (NLM CHEM, LitCovid) for evaluating biomedical NLP systems.

Purpose of the Study:

• To evaluate state-of-the-art biomedical-specific BERT models for the NLM CHEM and LitCovid tracks.
• To propose and validate a BERT-based ensemble learning approach for improved performance.
• To assess the effectiveness of a novel Medical Subject Headings identifier (MeSH ID) normalization algorithm.

Main Methods:

• Exploration and ensemble of various biomedical-specific pre-trained BERT models.
• Application of the ensemble approach to the NLM CHEM and LitCovid tracks of the BioCreative VII Challenge.
• Development and evaluation of a Medical Subject Headings identifier (MeSH ID) normalization algorithm for entity normalization.

Main Results:

• Achieved F1-scores of 85% (strict) and 91.8% (approximate) on the NLM-CHEM track.
• The MeSH ID normalization algorithm yielded F1-scores of approximately 80% in both strict and approximate evaluations.
• Demonstrated state-of-the-art performance on the LitCovid track for COVID-19 literature topic detection, outperforming existing methods.

Conclusions:

• The proposed BERT-based ensemble learning approach significantly improves performance in biomedical text analysis tasks.
• The MeSH ID normalization algorithm is effective for accurate entity normalization in biomedical contexts.
• The methodology establishes a new benchmark for topic detection in COVID-19 literature.