CLASSIFICATION OF PUBLIC OPINION ON FOREST MANAGEMENT IN SOUTHEAST SULAWESI PROVINCE BASED ON SOCIAL MEDIA COMMENTS USING AN LSTM-BASED DEEP LEARNING MODEL

Abstract

Forest management requires information support capable of rapidly capturing public responses. However, social media opinions are unstructured and difficult to analyze manually on a large scale. This study developed an opinion classification model based on Bidirectional Long Short-Term Memory (BiLSTM) to categorize YouTube comments related to forest management in Southeast Sulawesi Province into positive, negative, and neutral sentiment classes. A total of 1,875 comments were collected from 40 public videos across 30 YouTube channels using the YouTube Data API v3. Following initial labeling and preprocessing, 1,838 comments were retained for model development. The dataset was stratified into 1,286 training samples, 184 validation samples, and 368 testing samples. The model architecture consisted of a 128-dimensional embedding layer, 128 bidirectional LSTM units, dropout regularization, and a three-class softmax output layer. On the test dataset, the model achieved an accuracy of 89.40%, weighted precision of 89.49%, weighted recall of 89.40%, and weighted F1-score of 88.83%. The neutral class achieved an F1-score of 94.01%, outperforming the positive and negative classes, which obtained F1-scores of 64.71% and 58.82%, respectively. These findings indicate that the model is effective for overall sentiment classification; however, class imbalance limits its ability to consistently identify minority classes. Therefore, the classification results are more appropriately utilized as preliminary analytical support for understanding public opinion, while further improvements in labeling quality and data balancing are necessary to enhance classification performance.