News Classification with State-of-the-Art Deep Learning Methods

The subject of text classification is a fundamental area particularly considering current expansion of chatbots and recommendation systems. There exists promising results with traditional machine learning (ML) algorithms, on top of that, the deep learning methods proved superior performance with regard to accuracy and loss scores as evaluation metrics. This study focuses on increasing performance of text classification on AG News dataset with techniques of hyperparameter tuning on word-level convolutional neural network (CNN) model, bidirectional long-short term memory (BiLSTM) model and finally bidirectional encoder representations from transformers (BERT) architecture. The results denote outstanding performance of encoder based BERT transformer architecture on news classification.