Abstract

We sought to develop classifiers for automatically annotating neuroimaging research articles with labels from the Cognitive Paradigm Ontology. We examined classifier performance when varying corpus (abstract-only vs full-text), feature space (bag-of-words vs Cognitive Atlas), classifiers (BnB, knn, lr, svm). The most optimal classification performance across labels utilized full-text with bag-of-words and the logistic regression algorithm. When only considering abstract-text, the Cognitive Atlas features outperformed the bag-of-words features. We also found that anatomical terms dominated the features used for classification/