AI for AI: Using AI methods for classifying AI science documents

Abstract

Abstract Subject area classification is an important first phase in the entire process involved in bibliometrics. In this paper, we explore the possibility of using automated algorithms for classifying scientific papers related to Artificial Intelligence at the document level. The current process is semimanual and journal based, a realization that, we argue, opens up the potential for inaccuracies. To counter this, our proposed automated approach makes use of neural networks, specifically BERT. The classification accuracy of our model reaches 96.5%. In addition, the model was used for further classifying documents from 26 different subject areas from the Scopus database. Our findings indicate that a significant subset of existing Computer Science, Decision Science, and Mathematics publications could potentially be classified as AI-related. The same holds in particular cases in other science fields such as Medicine and Psychology or Arts and Humanities. The above indicate that in subject area classification processes, there is room for automatic approaches to be utilized in a complementary manner with traditional manual procedures.