X Sentiment Analysis on Indonesia’s New Capital (IKN) Using TF-IDF+SVM and IndoBERT, and Its Policy-Monitoring Implications

Abstract

This study investigates public perceptions of relocating Indonesia’s national capital to Ibu Kota Nusantara (IKN) through sentiment analysis of Indonesian-language X data, with two target classes: positive and negative. We compare two complementary modeling routes to balance semantic capacity and operational reliability. The lexical route pairs TF-IDF with a linear Support Vector Machine (SVM), providing a lightweight, stable, and reproducible baseline. The contextual route employs IndoBERT, a transformer model tailored to Indonesian, designed to capture implicit meaning and long-range dependencies within sentences. Preprocessing follows contemporary Indonesian NLP practice Unicode normalization, lowercasing, removal of URLs, mentions, and hashtags, normalization of slang into standard forms, removal of numerals and stopwords, and compression of elongated characters to stabilize lexical signal and reduce tokenization artifacts. Because the test data are imbalanced (the positive class is larger), evaluation emphasizes macro-F1 and negative-class recall so that overall accuracy is not inflated by the majority class. Final runs show TF-IDF+SVM achieves Accuracy 0.8908 and macro-F1 0.8709 with negative-recall 0.839; IndoBERT achieves Accuracy 0.9488 and macro-F1 0.9389 with negative-recall 0.920. The recall gain reduces undetected criticism and strengthens the practical value of a social-listening dashboard for governance and environmental issues where early warning is crucial.