Classification of Review Text using Hybrid Convolutional Neural Network and Gated Recurrent Unit Methods

Abstract

Consumer reviews are opinions from buyers to sellers based on service satisfaction or product quality. The more consumer reviews cause the process of analyzing manually will be difficult. Therefore, an automated sentiment analysis system is needed. Each review will be grouped into a sentiment class which is divided into positive and negative classes. This study aims to classify review texts using the Convolutional Neural Network (CNN) and Gated Recurrent Unit (GRU) methods. The research stages in this study include collecting data on Tokopedia review texts, extracting hidden information from review texts using CNN, conducting learning on review texts using GRU. A total of 1000 review texts were divided into 80% training data and 20% test data. The review text is converted into matrix using One Hot Encoding algorithm and then extracted using CNN. The CNN process includes the convolution calculation, the calculation of the Rectified Linear Unit (ReLU) activation function, and the pooling stage. The extraction results in the CNN process are continued in the GRU process. The GRU process includes initializing parameters, GRU feed forward, Cross-Entropy Error calculation, GRU feedback, and updating weights and biases. The optimal weight is obtained when the error value in the training is less than the expected minimum error or the training iteration has reached the specified maximum iteration. Optimal weight is used for validation test on test data. The implementation of review text classification using the hybrid Convolutional Neural Network (CNN) and Gated Recurrent Unit (GRU) method was made using the python programming language. The accuracy of the validation test is 88.5%