Hybrid Sentiment Analysis of Public Perception on Indonesia’s Role in the Board of Peace Using Inset Lexicon and Support Vector Machine

Abstract

Although sentiment analysis is increasingly utilised in Indonesian social media, there is a paucity of rigorous studies assessing hybrid lexicon–machine learning frameworks in the realm of foreign policy debate. This study fills this gap by presenting a hybrid sentiment analysis model that combines the domain-specific InSet Lexicon with Support Vector Machine (SVM) classification to analyse public perception of Indonesia’s involvement in the Board of Peace program—a United States-led international peace initiative in the Middle East region. This research utilises computational sentiment modelling in international peace diplomacy, a largely neglected area in Indonesian text mining, in contrast to prior studies that primarily concentrate on product reviews or local policy issues. A dataset comprising 1,454 Twitter (X) postings was collected and subjected to systematic preprocessing including case folding, tokenisation, slang normalisation, stopword removal, and Sastrawi-based stemming. The preprocessed data was automatically annotated utilising the InSet lexicon to produce pseudo-labels and subsequently classified employing SVM with two feature representation methodologies: TF-IDF and Word2Vec. Experimental findings indicate that the TF-IDF-based hybrid model attained the highest classification accuracy of 84% on the testing dataset, correctly predicting 1,221 out of 1,454 instances, surpassing the Word2Vec method (69%). Detailed per-class evaluation using precision, recall, and F1-score revealed strong negative-class performance (F1 = 0.91) while the neutral class remained most challenging due to class imbalance, with a macro-F1 of 0.70. A single 60:40 train-test split was applied; the pseudo-labelled nature of the dataset is acknowledged as a limitation requiring future manual validation. The results indicate that statistical term-weighting techniques are more resilient than semantic embedding representations in the context of domain-specific Indonesian policy discourse. This study methodologically contributes by empirically comparing feature representation options within a hybrid lexicon–SVM framework and substantively by offering computational evidence of polarised public attitude toward Indonesia’s diplomatic engagement. The findings underscore the significance of domain-specific lexicons in enhancing sentiment classification efficacy in low-resource language settings.