<![CDATA[The rapid growth of e-commerce in Indonesia has positioned customer reviews as a primary source of information reflecting consumer satisfaction and dissatisfaction with products. This study analyzes sentiment in Samsung product reviews on Tokopedia using natural language processing techniques. Data were collected from reviews on the official Samsung store on Tokopedia, followed by text pra-pemrosesan stages including cleaning, slang normalization, noise removal, selective stemming with the Sastrawi library, and stratified data splitting (80% training, 20% testing). Feature extraction was performed using Term Frequency–Inverse Document Frequency (TF-IDF) with n-gram integration (1-2), followed by training a Multinomial Naïve Bayes model optimized through GridSearchCV hyperparameter tuning. The primary objective was to develop an accurate sentiment classification model and identify dominant word patterns that reflect customer opinions. Evaluation results demonstrated an overall accuracy of 91.28%, with macro-average precision of 0.9112, recall of 0.9123, and F1-score of 0.9117. The negative class achieved the highest precision (0.9263), while the positive class showed strong recall (0.9079). TF-IDF analysis and word cloud visualization revealed that the words “barang” (product) and “kirim” (send) dominated across the dataset, with negative patterns centering on “lambat” (slow), “kecewa” (disappointed), and “rusak” (damaged), and positive patterns dominated by “bagus” (good), “cepat” (fast), “original”, and “mantap” (excellent). This study concludes that the Multinomial Naïve Bayes model is effective for sentiment analysis of unstructured Indonesian-language e-commerce reviews and provides valuable insights for official Samsung sellers to improve delivery services and product descriptions. Limitations in handling ambiguous or sarcastic reviews suggest opportunities for future research using transformer-based models such as IndoBERT or data augmentation techniques.]]>
