A Text Mining Pipeline Using Active and Deep Learning Aimed at Curating Information in Computational Neuroscience

Area of Science:

• Neuroscience
• Computational Neuroscience
• Neuroinformatics

Background:

• Manual curation of neuroscience entities is time-consuming and inefficient for large-scale brain modeling projects.
• Text mining offers a systematic solution for extracting relevant information from scientific literature.

Purpose of the Study:

• To apply text mining methods for efficient extraction of neuroscience entities.
• To develop and evaluate machine learning models for Named Entity Recognition (NER) in neuroscience.
• To create a methodology for training NER tools with minimal data for broad neuroscience applicability.

Main Methods:

• Two computational neuroscientists annotated a neuroscience corpus using active learning.
• Machine learning models, including a deep learning approach, were trained to recognize entities like Neuron Types, Brain Regions, and Experimental Values.
• Performance was compared against a traditional rule-based approach and a conditional random field model.

Main Results:

• The deep learning model demonstrated superior performance in identifying neuroscience entities.
• Achieved a macro average precision of 0.866, recall of 0.817, and F1 score of 0.837.
• Developed NER tools with performance comparable or superior to existing methods.

Conclusions:

• The proposed text mining methodology enables efficient and accurate extraction of neuroscience entities.
• The developed NER tools and training methodology require minimal data, making them adaptable to various neuroscience sub-domains.
• A valuable annotated corpus and guidelines are provided for reproducibility and further research.