Turtling: a time-aware neural topic model on NIH grant data.

Area of Science:

• Computational linguistics
• Bibliometrics
• Health services research

Background:

• Federal grant transparency initiatives necessitate advanced methods for knowledge extraction from large text volumes.
• Traditional survey methods are insufficient for analyzing multifaceted and evolving grant topics.
• Computational modeling of grant texts presents challenges due to their complexity and dynamic nature.

Purpose of the Study:

• To introduce Turtling, a time-aware neural topic model designed for analyzing federal grant texts.
• To improve the quality of extracted research topics and enable prediction of funding institutes.
• To provide insights into the evolution of research topics over time.

Main Methods:

• Utilizing pretrained biomedical word embeddings for research topic extraction.
• Employing a probabilistic time-series model for smooth and coherent topic evolution.
• Incorporating topic diversity loss and funding institute classification loss for enhanced topic quality and prediction.

Main Results:

• Turtling significantly outperforms existing methods on topic quality metrics when applied to NIH grant texts.
• The model successfully detects research topic trends across decades, offering insights into scientific evolution.
• Demonstrated effectiveness in improving topic quality and facilitating funding institute prediction.

Conclusions:

• Turtling is a valuable tool for analyzing federal grant texts, offering superior topic modeling capabilities.
• The model provides a dynamic view of research trends, aiding in understanding scientific progress.
• Turtling is available as open-source software, promoting wider adoption in grant text analysis.