Study of Web of Science Samples Using Neural Network Classifiers

The paper analyzes a text collection compiled from the Web of Science database. The study is aimed at assessing the quality of classification of scientific articles using modern neural network classifiers (all articles are given in the form of bibliographic descriptions, which include titles and abstracts). Despite the leadership of the Bidirectional Encoder Representations from Transformers neural network in many areas of natural language processing, its results on the Web of Science collection significantly depend on the “complexity” of the sample (the number of classes and their balance). In particular, on a “simple” sample (7 classes), the F1-score value was 0.908, and on a “complex” sample (134 classes), 0.815. The work also explored less resource-intensive neural network approaches. In particular, high quality indicators on the Web of Science collection were achieved using the FastText model for document representation and a convolutional neural network for classification (simple sample F1 = 0.888 and complex sample F1 = 0.766). At the same time, the combined use of FastText neural networks and a convolutional neural network can significantly reduce training time and classification speed compared to Bidirectional Encoder Representations from Transformers.