Automated annotation of scientific texts for ML-based keyphrase extraction and validation

Area of Science:

• Environmental genomics
• Microbiome science
• Bioinformatics

Background:

• Advanced omics technologies generate vast datasets lacking crucial metadata for effective search and utilization.
• Manual metadata curation and text labeling are time-consuming, hindering scientific progress.
• Environmental genomics and microbiome science fields require improved metadata strategies.

Purpose of the Study:

• To develop and present novel automated text labeling approaches for validating machine learning-generated metadata.
• To address the urgent need for efficient metadata annotation in data-rich scientific fields.
• To enhance the findability and usability of scientific datasets through automated annotation.

Main Methods:

• Developed two automated text labeling techniques for ML-generated metadata validation.
• Technique 1: Exploited relationships between diverse data sources (e.g., publications, proposals) for the same study.
• Technique 2: Leveraged domain-specific controlled vocabularies or ontologies for label generation.

Main Results:

• Proposed label assignment approaches generated both generic and highly specific text labels for unlabeled scientific texts.
• Achieved up to 44% label agreement with a machine learning keyword extraction algorithm.
• Demonstrated the potential of leveraging existing information to validate ML models for metadata extraction.

Conclusions:

• Automated text labeling offers a viable solution to the metadata challenge in omics research.
• The presented techniques effectively validate machine learning-derived metadata, improving data discoverability.
• These methods are particularly applicable to environmental genomics, accelerating data annotation and scientific innovation.