ANALISIS SENTIMEN MENGGUNAKAN MODEL CONCATENATION CLASSIFICATION INDOBERT UNTUK ULASAN PENGGUNA APLIKASI MYTELKOMSEL DI PLAYSTORE

Abstrak

In the digital era, user reviews of mobile applications have become an important data source for evaluating and improving service quality. This study aims to analyze user sentiment toward the MyTelkomsel application on Google PlayStore using the Concatenation Classification IndoBERT approach, which combines embeddings from IndoBERT and FastText to enrich the semantic representation of text. A total of 30,000 user reviews in Indonesian from 2022 to 2024 were collected through web scraping using the google-play-scraper library. The data was processed through several preprocessing stages, including normalization, cleaning, stopword removal, and stemming, followed by sentiment labeling into five categories: very negative, negative, neutral, positive, and very positive, based on polarity scores. Modeling was performed by combining the [CLS] token vector from IndoBERT (768 dimensions) and the FastText vector (300 dimensions), resulting in a 1068-dimensional vector. The dataset was split into 80% training data, 20% testing data and 80% training data, 10% testing data and 10% validation data . Evaluation using metrics such as accuracy, precision, recall, and F1-score showed high performance, with accuracy reaching 95% in 2022–2023 and increasing to 96% in 2024. These results indicate that the IndoBERT concatenation approach significantly improves sentiment classification accuracy and is effective in handling unstructured user review texts