Article-level classification of scientific publications: A comparison of deep learning, direct citation and bibliographic coupling

• DOI: 10.1371/journal.pone.0251493
• Publication Date: 2021-05-11T17:29:18Z
• Authors: Maxime Rivest; Etienne Vignola-G...; Éric Archambault
• Keywords: Bibliographic coupling; Citation analysis

Classification schemes for scientific activity and publications underpin a large swath of research evaluation practices at the organizational, governmental, national levels. Several classifications are currently in use, they require continuous work as new classification techniques becomes available topics emerge. Convolutional neural networks, subset “deep learning” approaches, have recently offered novel highly performant methods classifying voluminous corpora text. This article benchmarks deep learning technique on more than 40 million articles tens thousands scholarly journals. The comparison is performed against bibliographic coupling-, direct citation-, manual-based classifications—the established most widely used approaches field bibliometrics, by extension, many science innovation policy activities such grant competition management. results reveal that performance this first iteration approach equivalent to graph-based bibliometric approaches. All presented also par with manual classification. Somewhat surprisingly, no machine were found clearly outperform simple label propagation citation. In conclusion, promising because it just well other but has flexibility be further improved. For example, network incorporating information from citation likely hold key an even better algorithm.