HIERARCHICAL TEXT CLASSIFICATION FOR WEB OF SCIENCE SCIENTIFIC FIELDS

Abstract

This research focuses on making an efficient text classifier to map given corpora to specific scientific fields. Our study is set on the classification of different scientific fields according to the categories of the Web of Science (WOS). We design and develop various deep learning architectures such as Convolutional Neural Network (CNN), Deep Neural Network (DNN), and Recurrent Neural Network (RNN) at both the parent and child levels. To make our model perform better, we effectively use Hyperband Tuning. We aim to construct a precise hierarchical text classifier for lower levels, and smaller general model sizes. The evaluation employs a special metric known as the hierarchical confusion matrix. Results based on a broad investigation of Word Embedding, Document Embedding, and Hyperband Tuning show that the hierarchical combination of CNN and DNN in parent-child levels can achieve greater accuracy. Our model scored genuinely well, with an F1 score of 94.29% and an accuracy of 99.33%. Although using an RNN at the parent level and another at the child level led to lower accuracy, it effectively reduced the overall model size. We also conducted a comprehensive evaluation of various model architectures using the AoI2WoS dataset. By incorporating Google News word embeddings, we tested different combinations of RNN-DNN and RNN-RNN models on the AoI2WoS dataset. The RNN-DNN model yielded the best results, achieving an accuracy of 98.71% and an F1-score of 91.87%. These findings not only push forward the development of hierarchical text classification but also provide potent tools for utilizing research in scientometrics, and bibliometric researches.