Hoax Classification in Indonesian Language with Bidirectional Temporal Convolutional Network Architecture

Abstract

The increasingly massive rate of information dissemination in cyberspace has had several negative impacts, one of which is the increased vulnerability to the spread of hoaxes. Hoax has seven classifications. Classification problems such as hoax classification can be automated using the application of the Deep Learning model. Bidirectional Temporal Convolutional Network (Bi-TCN) is a type of Deep Learning architectural model that is very suitable for text classification cases. Because the architecture uses dilation factors in its feature extraction so it can generate exceptionally large receptive fields and is supported by Bidirectional aggregation to ensure that the model can learn long-term dependencies without storing duplicate context information. The purpose of this study is to evaluate the performance of Bi-TCN architecture combined with pre-trained FastText embedding model for hoax classification in Indonesian and implement the resulting model on website. Based on the research that has been done, the model with Bi-TCN architecture has satisfactory performance with an accuracy score of 92.98% and a loss value that can be reduced to 0.191. Out of a total of 13,673 data tested with this model, only 414 data or in other words around 3% of the total data were incorrect predictions.