A Hybrid Rule-Based and Multinomial Naïve Bayes System for Sentiment and Intent Classification of Indonesian Public Reports with Sarcasm Detection

Abstrak

Public service reporting systems in Indonesia face significant challenges in processing large volumes of unstructured citizen feedback efficiently. This study proposes Aspiralytica, a mobile-based citizen report classification system that integrates TF-IDF feature extraction with a Multinomial Naive Bayes (MNB) classifier within a hybrid rule-based and machine learning architecture. The system simultaneously performs three-class sentiment classification (positive, negative, neutral) and five-class intent classification (complaint, appreciation, request, emergency, suggestion), with automated priority level determination and a rule-based sarcasm detection module achieving F1 of 0.8980. Evaluated on an augmented dataset of 1,137 sentiment-labeled and 2,187 intent-labeled Indonesian-language citizen report texts using Stratified 10-Fold Cross-Validation, the proposed MNB model achieved sentiment classification accuracy of 96.59% (F1: 96.59%) and intent classification accuracy of 96.97% (F1: 96.96%). An ablation study confirmed TF-IDF with MNB as the dominant performance driver, and a computational efficiency benchmark empirically justified MNB selection with mean inference latency of 0.3955 ms and throughput of 52,312 requests per second. The system is deployed as a FastAPI backend integrated with a React Native mobile frontend, delivering real-time classification through a citizen-facing interface.

Identifier

• OpenAlex: https://openalex.org/W7201834274
• DOI: https://doi.org/10.24076/intechnojournal.2026v8i1.2765