Analisis Sentimen Berbasis Topik pada Ulasan Aplikasi Maxim Menggunakan Pendekatan Semisupervised Learning dengan Indobert, Lda, dan Generative Ai Laporan Skripsi

Abstract

Maxim is a ride-hailing service that has been operating in Indonesia since July 2018. The number of users continues to grow, which also increases the number of reviews on platforms like Google Play Store and X (Twitter). These reviews contain a lot of information about user perceptions of the service. However, the large volume makes manual analysis difficult and impractical to do. This study uses sentiment analysis to understand user perceptions of Maxim. The approach combines IndoBERT with semi-supervised learning for sentiment classification, Latent Dirichlet Allocation (LDA) for topic extraction, and Generative AI for topic interpretation. Data was collected from 2018 to 2025, totaling 252.321 reviews. After preprocessing, 209,665 reviews remained usable. An initial dataset of 2,500 reviews was manually labeled by three annotators. Inter-rater agreement was tested using Fleiss' Kappa and obtained a score of 0.6286, which falls in the substantial agreement category. This labeled dataset was used to train IndoBERT-base-p2 as the initial model. Iterative bootstrapping was then carried out for five iterations by adding pseudo-labels with a confidence score of at least 0.98. The best model, M3, reached 96% accuracy and a macro F1-score of 0.93. For topic modeling, LDA identified five main topics: Driver Attitude, Application Performance, Order Process, GPS/Map Accuracy, and Price and Fare. Topic labeling was done using Google Gemini through a structured prompt. Results show that Driver Attitude is mostly positive, while Application Performance, Order Process, and GPS/Map Accuracy are dominated by negative sentiment. Price and Fare has a relatively balanced distribution. Sentiment trends from 2018 to 2024 show gradual improvement. The findings suggest that technical issues in the application and the driver-matching process are the main areas that need improvement.