Full-text chemical identification with improved generalizability and tagging consistency

Area of Science:

• Biomedical Informatics
• Computational Chemistry
• Natural Language Processing

Background:

• Chemical identification in scientific literature is crucial for data mining and knowledge discovery.
• Existing models for chemical named entity recognition and normalization are primarily evaluated on titles and abstracts, with limited validation on full-text articles.
• Full-text analysis reveals limitations in current models, including poor generalizability to novel mentions and inconsistent tagging.

Purpose of the Study:

• To address the limitations of existing chemical identification models in full-text scientific articles.
• To develop and evaluate improved methods for chemical named entity recognition and normalization.
• To enhance the accuracy and reliability of extracting chemical information from full-text scientific documents.

Main Methods:

• Implemented transfer learning and mention-wise majority voting to improve generalizability and consistency in named entity recognition.
• Developed a hybrid model combining neural and dictionary-based approaches for named entity normalization, balancing recall and precision.
• Utilized simple training and post-processing techniques to refine model performance.

Main Results:

• Achieved first place in the BioCreative VII NLM-Chem challenge for named entity recognition with an 86.72 F1 score.
• Obtained an 78.31 F1 score for named entity normalization in the challenge, outperforming the median.
• Post-challenge re-implementation yielded an 84.70 F1 score for normalization, surpassing the challenge's best score by 3.34 F1 points.

Conclusions:

• The proposed methods effectively address limitations in chemical identification within full-text articles.
• The hybrid model demonstrates superior performance in both named entity recognition and normalization tasks.
• This work advances the state-of-the-art in automated chemical information extraction from scientific literature.