Sentiment Analysis of Traveloka App User Reviews Using Word2vec and LSTM

Abstract

Objective: As the travel trend in Indonesia increases, online travel agent (OTA) services such as Traveloka are becoming increasingly popular. However, with high competition in this industry, companies need to understand customer sentiment to improve service quality. This study aims to develop an automated sentiment analysis model on Traveloka app user reviews using a deep learning approach with Long 0Short-Term Memory (LSTM) and Word2Vec word representation. Design/method/approach: This study uses a quantitative method with stages including data collection, preprocessing, labeling, and modeling. The data used comes from Kaggle, which contains 12,689 Traveloka user reviews on the Google Play Store between January and October 2023. Preprocessing is carried out using case folding, tokenization, stopword removal, and stemming. Next, the data is represented in vector form using Word2Vec before being trained with an LSTM model. Experiments are conducted with various vector sizes (100, 200, 300, and 400) to evaluate their effect on model accuracy. Results: The test results show that the LSTM model with a vector size of 400 and a training dataset of 4073 achieved the highest accuracy of 89%, while vector sizes of 100, 200, and 300 produced accuracies of 85%, 86%, and 82%, respectively. This indicates that the dimensionality of word representation and the amount of data can affect the model's performance in understanding user sentiment. Originality/state of the art: This study contributes to the development of deep learning-based sentiment analysis models in Indonesian, especially in the travel services sector. Different from previous studies that mostly use shallow learning methods such as Naïve Bayes and SVM, this study adopts the Word2Vec and LSTM approaches that are more effective in capturing semantic relationships between words. This study also provides insights into the effect of Word2Vec vector size and training dataset size on LSTM in improving the accuracy of sentiment analysis models.