Klasifikasi Sentimen Berbasis GPT dan Pemodelan Topik LDA terhadap Opini Publik pada Platform X mengenai Isu Program Makan Bergizi Gratis

Abstract

Penelitian ini bertujuan untuk menganalisis persepsi publik terhadap Program Makan Bergizi Gratis (MBG) di Indonesia melalui pemodelan topik dan analisis sentimen terhadap data media sosial X. Sebanyak 7.004 tweet dikumpulkan dan diproses melalui beberapa tahapan preprocessing, yaitu filtering, pembersihan teks, normalisasi, tokenisasi, stopword removal, dan stemming. Pemodelan topik dilakukan menggunakan metode Latent Dirichlet Allocation (LDA), sedangkan klasifikasi sentimen dilakukan menggunakan model Generative Pre-trained Transformer (GPT). Hasil penelitian menunjukkan bahwa model LDA optimal menghasilkan empat topik utama, yaitu (1) Pelaksanaan dan Kontroversi Program MBG, (2) Kebijakan Pangan dan Dinamika Politik, (3) Pengembangan Sumber Daya Manusia dan Prioritas Anggaran, serta (4) Isu Pendukung dan Diskursus Kontekstual. Di antara keempat topik tersebut, pembahasan mengenai pelaksanaan program dan berbagai kontroversi yang menyertainya merupakan topik yang paling dominan. Model GPT menunjukkan kinerja klasifikasi sentimen yang baik, meskipun masih memiliki keterbatasan dalam mengklasifikasikan sentimen netral. Model ini memperoleh nilai akurasi sebesar 84,01%, balanced accuracy sebesar 72,28%, weighted F1-score sebesar 84,94%, dan macro F1-score sebesar 68,92%. Analisis sentimen berdasarkan waktu dan topik menunjukkan bahwa sentimen negatif secara konsisten mendominasi diskusi publik pada platform X selama periode pengamatan maupun pada seluruh topik yang diidentifikasi. Temuan tersebut mengindikasikan bahwa persepsi publik di X terhadap Program MBG masih didominasi oleh kritik, kekhawatiran, dan ketidakpuasan, terutama terkait pelaksanaan program serta kebijakan-kebijakan yang mendukungnya.

This study aims to analyze public perceptions of the Free Nutritious Meals Program (MBG) in Indonesia through topic modeling and sentiment analysis of social media data from X. A total of 7,004 tweets were collected and processed through several preprocessing steps, namely filtering, text cleaning, normalization, tokenization, stopword removal, and stemming. Topic modeling was performed using the Latent Dirichlet Allocation (LDA) method, while sentiment classification was performed using the Generative Pre-trained Transformer (GPT) model. The results of the study show that the optimal LDA model yielded four main topics, namely (1) Implementation and Controversies of the MBG Program, (2) Food Policy and Political Dynamics, (3) Human Resource Development and Budget Priorities, and (4) Supporting Issues and Contextual Discourse. Among these four topics, discussions regarding the program’s implementation and the various controversies surrounding it were the most dominant. The GPT model demonstrated good sentiment classification performance, although it still has limitations in classifying neutral sentiments. This model achieved an accuracy of 84.01%, a balanced accuracy of 72.28%, a weighted F1-score of 84.94%, and a macro F1-score of 68.92%. Sentiment analysis based on time and topic shows that negative sentiment consistently dominated public discussions during the observation period and across all identified topics. These findings indicate that public perception on X regarding the MBG Program is still dominated by criticism, concerns, and dissatisfaction, particularly regarding the program’s implementation and the policies supporting it.