The Best Parameter Tuning on RNN Layers for Indonesian Text Classification

Abstract

Recurrent Neural Network (RNN) is a deep learning architecture commonly used to process time series and sequence data. Various architectures have been developed to improve the performance of the algorithm in terms of both accuracy and computation time. Besides, the use of appropriate parameter values when building a neural network model also plays an important role in the quality and the outcome of the learning model. In this study, the model trained using RNN-Vanilla, LSTM, and GRU each with 4 different combinations of parameter settings, namely bidirectional mode (True, False), the number of neuron units on each layer (64, 128, 256), the number of RNN layers on the neural network (1, 2, 3), and the batch size when training the model (32, 64, 128). By combining all the parameter values, 162 trials were carried out to perform the task of classifying Indonesian language customer support tickets with four category classes. This study gives the result that the same network architecture but with different parameter combinations results in significant differences in the level of accuracy. The lowest accuracy of all experiments was 32.874% and the highest accuracy resulted was 84.369%. Overall, by calculating the average accuracy of each parameter value, the results obtained are: GRU has the best performance, accuracy tends to increase by activating bidirectional mode, increasing the number of neuron units in the hidden layer, and reducing the batch size. Meanwhile, the addition of the number of RNN layers on the neural network has no impact on increasing the level of accuracy.