Analisis Sentimen Ulasan Produk Kuliner Lokal Menggunakan Model BERT: Studi Kasus pada UMKM Pawonkoe

Abstract

This research investigates the application of the Bidirectional Encoder Representations from Transformers (BERT) model for sentiment analysis on customer reviews of Pawonkoe’s local food products. Customer feedback was collected through an online questionnaire containing textual opinions and multi-aspect product ratings. To ensure high-quality model input, the dataset underwent systematic preprocessing, including text cleaning, normalization, tokenization, and handling of informal Indonesian expressions. Smart data re-labeling techniques and sentiment scoring were used to reduce labeling inconsistencies and address class imbalance. A multilingual BERT model was fine-tuned to classify sentiments into positive, neutral, and negative categories. Experimental results demonstrated a steady improvement in performance, achieving 73.97% accuracy and a weighted F1-score of 0.74. The model effectively identified positive and neutral sentiments, although negative sentiment detection remained challenging due to limited samples. Overall findings indicate that customer perceptions toward Pawonkoe products are predominantly positive, particularly regarding taste, texture, and overall quality. This study highlights the effectiveness of transformer-based NLP models in extracting actionable insights from customer opinions and provides valuable recommendations for SMEs to strengthen product quality and digital marketing strategies.