Integrating curation into scientific publishing to train AI models

Area of Science:

• Biomedical Informatics
• Computational Biology
• Scientific Publishing

Background:

• High-throughput data extraction and structured labeling from academic articles are vital for machine learning and secondary analyses.
• Existing methods lack integration with the publishing workflow and comprehensive annotation of experimental roles and methodologies.
• There is a need for advanced bioentity recognition and annotation within the scientific literature.

Purpose of the Study:

• To embed multimodal data curation into the academic publishing process.
• To create a comprehensive dataset for training AI models in biomedical research.
• To improve the annotation accuracy of figure panels and captions.

Main Methods:

• Integrated multimodal data curation into the academic publishing workflow.
• Utilized natural language processing and author feedback for annotation.
• Annotated segmented figure panels and captions from molecular and cell biology articles.
• Developed new AI tasks for evaluating dataset utility, including named-entity recognition and context-dependent semantic analysis.

Main Results:

• Created the SourceData-NLP dataset with over 620,000 annotated biomedical entities from 18,689 figures across 3,223 articles.
• Annotations include eight bioentity classes and experimental roles/methodologies.
• Demonstrated the dataset's utility for AI model training in named-entity recognition, figure caption segmentation, and novel semantic tasks.
• Showcased multi-modal applications for segmenting figures into panels and captions.

Conclusions:

• The SourceData-NLP dataset significantly enhances machine learning applications in biomedical research.
• Integrating data curation into publishing streamlines the creation of valuable, structured datasets.
• The developed models and dataset facilitate advanced analysis of scientific figures and text.