<![CDATA[Existing disaster sentiment analysis mainly focuses on emotional polarity classification, while often over-looking the causal reasoning that shapes public discourse on responsibility for disaster outcomes. This study proposes and assesses a Long Short-Term Memory (LSTM)-based causal attribution classification framework to examine YouTube comments related to the 2025 Sumatra flash flood. It compares LSTM performance with Sup-port Vector Machine (SVM) and Naïve Bayes baselines. A total of 17,503 publicly available comments were collected through the YouTube Data API v3 and processed into a final dataset of 12,299 comments. The com-ments were classified into two causal categories, human factor and nature/prayer factor, using lexicon-based scoring validated by three independent annotators (Cohen's κ = 0.81). The experimental results show that LSTM achieves 98.17% accuracy with strong stability (±0.25% standard deviation) under stratified five-fold cross-validation, substantially outperforming SVM (82.83%) and Naïve Bayes (75.04%). These findings indi-cate that sequence-based architectures can capture the contextual dependencies in causal attribution dis-course, offering a replicable framework for disaster risk communication monitoring systems.]]>
Copyrights © 2026
