Sistem Pendeteksi Spam Pesan Singkat Bahasa Indonesia Menggunakan IndoBERT dan Multi-Graph Convolutional Network

Abstract

Short Message Service (SMS) in Indonesia remains a primary medium for disseminating unwanted spam content, ranging from aggressive commercial advertisements to harmful financial fraud attempts that pose significant risks to users. The core challenge in SMS spam detection lies in the dynamic linguistic complexity of the Indonesian language, which frequently incorporates slang, inconsistent abbreviations, and code- switching patterns that hinder traditional keyword-based filtering techniques. While transformer-based language models such as IndoBERT have demonstrated exceptional performance in semantic context understanding, they often struggle with structural text obfuscation designed to bypass automated security systems. This research proposes an innovative hybrid architecture that integrates the rich semantic representations of IndoBERT with the structural analysis capabilities of a Multi-Graph Convolutional Network (GCN). Through this approach, relational patterns between words are modeled as graph representations to augment the contextual depth of the base model. This study utilizes a curated dataset of Indonesian SMS messages to train and evaluate the performance of the proposed hybrid model. Experimental results demonstrate outstanding performance, with the model achieving a peak accuracy of 99.93% and an F1-Score of 1.00. Ablation studies further confirm that incorporating graph-based features significantly enhances detection precision and effectively minimizes false positive rates compared to using the standalone language model. The findings of this research provide significant theoretical contributions to the field of Natural Language Processing and are practically implemented as a real-time detection prototype to validate its robustness and feasibility in addressing spam threats within real-world communication scenarios.