Hierarchical Text Classification using CNNs with Local Classification Per Parent Node Approach

In this paper, authors present the application of Convolutional Neural Networks (CNNs) for Hierarchical Text Classification (HTC) using Local Classification per Parent Node (LCPN) approach. A 20Newsgroup hierarchical training dataset with 3 hierarchical levels and more than 20 categories was used for training, evaluation and testing of the model. Several variations of hyperparameters settings are involved in the experiments such as batch sizes, embedding sizes, and number of training samples. The comparison of achieved results is performed against baseline and hierarchical models such as Support Vector Machine (SVM), Logistic Regression (LR), flat CNN and CNN with Local Classification per Node (LCN) approach. The results indicated that HTC using CNNs with LCPN approach outperformed the rest of selected methods by a margin between 5% - 15%.