<![CDATA[Automated text moderation systems on public service platforms are often exploited by manipulative spam messages from brokers offering illegal financial services. Previous text classification studies have frequently prioritized high accuracy metrics while overlooking the impact of data leakage caused by repetitive spam templates, a methodological flaw that can lead to severe model overfitting. This study aims to design and optimize a Natural Language Processing (NLP) model using the IndoBERT-Lite architecture to distinguish between organic user complaints and manipulative broker-generated comments. The proposed methodology focuses on extreme data deduplication, reducing 55,156 raw records into a balanced dataset containing 4,626 unique samples (57.1% organic and 42.9% spam). The training process was optimized using Gradient Accumulation and Early Stopping to ensure genuine model generalization capability. The evaluation results demonstrate that the optimized model successfully mitigated the initial overfitting problem, achieving both accuracy and F1-score values of 98% on unseen test data. These findings provide a reliable and data leakage–free automated moderation solution for internal digital customer service systems.]]>
Copyrights © 2026
