Improving Nepali News Recommendation using Classification based on LSTM Recurrent Neural Networks

Abstract

News classification is the process of grouping news documents into some predefined categories. Due to the increasing volume of the Nepali news content being generated everyday by thousands of online news portals, appropriate classification of these news items has become a necessity for the news readers. This thesis was targeted to improve the Nepali news classification based on Recurrent Neural Networks, that uses deep layers of neural networks to classify the news to appropriate category. In this research paper, five popular news portals website across eight different categories was used for the purpose of data gathering and their classification accuracies was compared among these websites as well as overall accuracy was measured. The model was compared with the Support Vector Machine based on the parameters Accuracy, Precision, Recall and F1 Score. The use of Long Short Term Memory Recurrent Neural Network has improved the precision with the use of word2vec model. The presented model in the paper have achieved a good accuracy of 84.63% and precision of 89% in compared to the SVM where the accuracy was 81.41% and precision 85% Among five news websites compared onlinekhabar.com was found to have good classification of the
news where as ratopati.com was the least. Based on the categories of the news, sports news was classified more accurately by the model and economy was least accurately classified.