Analisis Sentimen Berbasis Aspek terhadap Pilkada Jember 2024 Menggunakan Model IndoBERT dan Algoritma LDA

Abstract

The political dynamics leading up to the 2024 Jember Regional Election (Pilkada) triggered a significant surge of public opinion on YouTube and TikTok platforms. The informal nature of social media language, including slang, abbreviations, and mixed local dialects, presents challenges for manual analysis. This study applies a Natural Language Processing (NLP) approach, beginning with preprocessing stages such as text cleaning, case folding, normalization, tokenization, stopword removal, and stemming to improve data quality. Topic modeling was conducted using Latent Dirichlet Allocation (LDA) to identify strategic issues discussed online, which were further refined using BERT Summarization to extract the core themes of each topic. Sentiment classification was performed using an IndoBERT model that had been fine-tuned on 6,500 samples from the IndoNLU SmSA dataset through stratified sampling. The analysis of 33,392 comments successfully identified nine major topics, with Leadership Image and Achievements emerging as the most dominant in terms of volume (5,719 comments). The topics with the highest proportion of negative sentiment were “Promises vs. Evidence” (68.9%) and Debate Performance Evaluation (60.3%). Overall sentiment distribution was dominated by negative sentiment (approximately 46%), followed by positive (32%) and neutral (21%). The IndoBERT model demonstrated strong performance, achieving 91.20% Accuracy, 91.23% Precision, 91.20% Recall, and a 90.94% F1-Score. These findings indicate that online discourse surrounding the 2024 Jember election was characterized more by critical evaluation than pure support, reflecting a significant polarization of public opinion in the digital political arena.