Weakly Supervised Sentiment Analysis of Indonesian Rural Tourism Reviews: A TF-IDF Baseline for Melung Tourism Village

Abstract

This study investigates sentiment classification of Indonesian-language tourist reviews from the rural destination of Melung Tourism Village. A total of 724 user-generated reviews from 546 unique users are preprocessed using Indonesian-specific text cleaning, stopword filtering, and stemming, then weakly labeled through a stemmed positive–negative lexicon. TF-IDF unigram–bigram features are extracted from the preprocessed texts and used to train three classical classifiers: Naive Bayes, linear Support Vector Machine (SVM), and Logistic Regression. To address class imbalance, RandomOverSampler is applied only to the training data, and model evaluation combines stratified 5-fold cross-validation with a held-out test set, using weighted F1-score as the primary metric. Logistic Regression achieves the best performance on the test set (weighted F1 = 0.8799, accuracy = 0.8828), closely followed by SVM, while Naive Bayes lags behind. The results show that, even with a modest, weakly supervised dataset, a carefully designed classical pipeline can yield reliable sentiment indicators to support data-driven management of rural tourism destinations.