Comparison of Classification for Indonesian Language News Documents Using Recurrent Neural Network (RNN) and Long Short Term Memory (LSTM) Algorithms

Abstrak

The development of online news has grown very fast. The high volume of text documents was triggered by activities from various news sources. Due to the large amount of news that is included on the website, sometimes the news is posted not according to its category which is most likely caused by human error. The grouping of online news is important for user convenience in searching for news according to its category. It need an intelligent system that can classify online news automatically. This research evaluates deep learning techniques using LSTM and RNN, and compared with the results obtained from previous studies, which used the NBC algorithm. To experiment the system, an Indonesia News Corpus with 7 different categories and total 2100 documents, collected by crawling online national news portals, is used. Due to the unbalanced number of class compositions or news categories, integration is also carried out SMOTE. The average empirical results show that the classification accuracy from RNN with SMOTE with an accuracy of 95.2% and followed by LSTM with SMOTE is 97.8%, both of which are able to outperform the NBC method with an accuracy of 73.2%.