Toward Computer-Assisted Text Curation: Classification Is Easy (Choosing Training Data Can Be Hard...)

摘要

We aim to design a system for classifying scientific articles based on the presence of protein characterization experiments, intending to aid the curators populating JCVI's Characterized Protein (CHAR) Database of experimentally characterized protein s. We trained two classifiers using small datasets labeled by CHAR curators, and another classifier based on a much larger dataset using annotations from public databases. Performance varied greatly, in ways we did not anticipate. We describe the datasets, the classification method, and discuss the unexpected results.