Implementasi Sistem Deteksi Sentimen Berbasis Aspek untuk Ulasan Smartphone Menggunakan Metode Support Vector Machine (Studi kasus: iPhone 17 Pro)

Abstract

Indonesian internet users has generated millions of smartphone reviews on YouTube, predominantly written in highly informal language containing abbreviations, typos, and sarcasm, making automated analysis considerably challenging for conventional systems. This research aims to develop an Aspect-Based Sentiment Analysis (ABSA) using Support Vector Machine and TF-IDF to classify iPhone 17 Pro reviews based on four product aspects, namely Camera, Battery, Price, and Design, across three sentiment classes: Positive, Negative, and Neutral. A seven-stage text preprocessing pipeline was designed, comprising case folding and data cleaning, typo and abbreviation normalization using a dictionary of over 200 entries, lexicon-based sarcasm handling, slang normalization, custom stopword removal, Sastrawi library stemming, and post-stemming negation handling to preserve negation token consistency. The dataset consisted of 658 YouTube comments divided using an 80:20 ratio (526 training and 132 testing data). Class imbalance in sentiment data was addressed using the SMOTE method. Evaluation results show the aspect classification model achieved an accuracy of 90.15 percent with a Weighted F1-Score of 0.90, while the sentiment classification model reached an accuracy of 82.58 percent with a Weighted F1-Score of 0.82, yielding an overall average system accuracy of 86.36 percent. The system was implemented as a Flask Python-based web application named GadgetSense, featuring automatic multi-clause analysis, a Blackbox Terminal for preprocessing transparency, and YouTube Data API v3 integration. The research demonstrates that SVM combined with a comprehensive text normalization pipeline is effective for ABSA on informal Indonesian language text.