Automatically Determining Versions of Scholarly Articles

Abstract

Background: Repositories of scholarly articles should provide authoritative information about the materials they distribute and should distribute those materials in keeping with pertinent laws. To do so, it is important to have accurate information about the versions of articles in a collection. Analysis: This article presents a simple statistical model to classify articles as author manuscripts or versions of record, with parameters trained on a collection of articles that have been hand-annotated for version. The algorithm achieves about 94 percent accuracy on average (cross-validated). Conclusion and implications: The average pairwise annotator agreement among a group of experts was 94 percent, showing that the method developed in this article displays performance competitive with human experts.