Chemical Entity Recognition for MEDLINE Indexing

Area of Science:

• Biomedical Informatics
• Natural Language Processing
• Scientific Literature Analysis

Background:

• Chemical entity recognition is crucial for indexing scientific literature in MEDLINE.
• The current Medical Text Indexer tool was adapted for chemical recognition, not originally designed for it.
• Existing methods rely on MetaMap and custom rules, prompting a need for improved tools.

Purpose of the Study:

• To develop and evaluate a more suitable tool for chemical entity recognition.
• To assess the performance of eleven chemical recognition systems on a curated dataset.
• To identify a system effective for both MEDLINE indexing and general chemical entity recognition.

Main Methods:

• Created a dataset of 200 MEDLINE titles and abstracts annotated with chemical entities.
• Evaluated eleven distinct chemical entity recognition systems using this dataset.
• Utilized a SciBERT ensemble as one of the evaluated systems.

Main Results:

• The SciBERT ensemble achieved the highest performance among the evaluated systems.
• The study identified a system effective for both indexing and broader chemical recognition tasks.
• Performance metrics indicated superior accuracy with the SciBERT ensemble.

Conclusions:

• The SciBERT ensemble shows significant promise for advancing chemical entity recognition in scientific literature.
• This research provides a benchmark for evaluating chemical recognition tools.
• Improved chemical entity recognition can enhance the indexing and retrieval of scientific information.