Statistical algorithms for ontology-based annotation of scientific literature

Area of Science:

• Bioinformatics
• Computational Linguistics
• Neuroscience

Background:

• Ontologies provide structured knowledge for data annotation, aiding search and meta-analysis.
• Manual annotation is time-consuming; current text mining lacks ontology structure utilization.
• Existing methods rely on keyword matching, missing deeper semantic relationships.

Purpose of the Study:

• To develop a probabilistic framework for automatic literature annotation using ontology structure.
• To enhance the annotation of human functional neuroimaging literature with the Cognitive Paradigm Ontology (CogPO).

Main Methods:

• A probabilistic framework models restrictions among ontology classes.
• Combines Naïve Bayes with decision trees for annotation.
• The framework is adaptable to evolving domain knowledge.

Main Results:

• The framework was evaluated against Naïve Bayes, Bayesian Decision Trees, and Constrained Decision Tree classifiers.
• Performance was measured using the F1-micro score, comparing different classification approaches.
• Human expert input was integrated into the Constrained Decision Tree models.

Conclusions:

• The proposed framework effectively models ontology knowledge, including dependencies and restrictions.
• It surpasses traditional text mining by leveraging implicit and explicit information within ontologies.
• Successful application demonstrated in annotating neuroimaging literature using CogPO.