<![CDATA[Community proposals in the form of regional development proposals are often voluminous, creating administrative barriers to verification and alignment with dynamic development priorities. This study aims to implement and evaluate an automatic classification system to help verifiers quickly and accurately prioritize community proposals that are most relevant to the regional development policy directions. This study employs a comparative experimental design to evaluate the performance of a Large Language Model (LLM) against traditional machine learning models, namely Multinomial Naive Bayes (MNB) and Complementary Naive Bayes (CNB). PDF documents in the dataset were processed to extract text, and Optical Character Recognition (OCR) was used for scanned documents. Documents were sorted based on seven administrative criteria, then summarized to extract key points from those meeting the criteria, classified according to development priorities, and evaluated using Stratified 10-Fold Cross-Validation to ensure consistent and representative results. Experimental results show that Complement Naive Bayes (CNB) achieved the highest performance with an accuracy of 87.69% and an F1-score of 0.8760, followed by LLM with an accuracy of 87.72% and an F1-score of 0.8759, and MNB with an accuracy of 84.3% and an F1-score of 0.8430. Although CNB has a slight edge in statistical accuracy, LLM offers added value through its semantic evaluation capabilities, which can provide the logical reasoning behind each classification decision. This study provides empirical evidence that lightweight traditional algorithms (CNB) remain more robust and computationally efficient than zero-shot LLMs for localized administrative texts, offering a scalable solution for government digital transformation.]]>
