Klasifikasi Sentimen Publik terhadap Program MBG di Media Sosial X dengan IndoBERTweet

Abstract

Abstract. The Free Nutritious Food Program (MBG) is Indonesia's national school feeding program, developed in line with the World Food Programme (WFP) framework. Since its launch on January 6, 2025, the program has generated diverse public responses, one of which has been expressed through social media platform X. The high volume of MBG-related tweets makes manual sentiment analysis inefficient, necessitating a computational approach based on Natural Language Processing (NLP). This study applies IndoBERTweet, a pre-trained language model based on the BERT architecture developed specifically for processing Indonesian Twitter text, to classify public sentiment regarding the MBG program. A total of 3,372 tweets were collected using web scraping with keyword "MBG" in December 2025. Data underwent preprocessing, manual sentiment labeling by three annotators using majority voting, tokenization, training, validation, testing data splitting, and fine-tuning using the IndoBERTweet model. After preprocessing, sentiment distribution consisted of 1,332 negative, 927 neutral, and 497 positive tweets from 2,756 total tweets. Evaluation results on testing data yielded accuracy of 73.67%, weighted precision of 73.73%, weighted recall of 73.67%, and weighted F1-score of 73.35%, categorized as "Good" performance. The negative sentiment class achieved the best performance with an F1-score of 80.19%, while neutral class had the lowest F1-score of 64.94%.

Abstrak. Program Makan Bergizi Gratis (MBG) merupakan program pemberian makan nasional di Indonesia yang mengacu pada kerangka World Food Programme (WFP). Sejak diluncurkan pada 6 Januari 2025, berbagai respons masyarakat terhadap program tersebut muncul di berbagai media, salah satunya melalui media sosial X. Tingginya volume cuitan terkait MBG menyebabkan analisis sentimen secara manual menjadi kurang efisien sehingga diperlukan pendekatan komputasional berbasis Natural Language Processing (NLP). Penelitian ini menerapkan IndoBERTweet, yaitu model bahasa pra-latih berbasis arsitektur BERT yang dikembangkan khusus untuk memproses teks Twitter berbahasa Indonesia, guna mengklasifikasikan sentimen publik terkait program MBG. Sebanyak 3.372 cuitan dikumpulkan menggunakan teknik web scraping dengan kata kunci "MBG" pada Desember 2025. Data melalui tahap preprocessing, pelabelan sentimen manual oleh tiga anotator dengan majority voting, tokenisasi, pembagian data training, validation dan testing, serta fine-tuning model IndoBERTweet. Distribusi sentimen setelah di preprocessing terdiri atas 1.332 tweet negatif, 927 tweet netral, dan 497 tweet positif dari total 2.756 tweet. Hasil evaluasi pada data testing menghasilkan accuracy sebesar 73,67%, weighted precision 73,73%, weighted recall 73,67%, dan weighted F1-score 73,35% yang termasuk kategori "Baik". Kelas sentimen negatif menghasilkan performa terbaik dengan F1-score 80,19%, sedangkan kelas netral memperoleh F1-score terendah sebesar 64,94%.