On online platforms, harmful contents such as hate speech, spam, and personal information leakage are key factors that negatively affect user experience. This study proposes a Two-Tier Hybrid Moderation Pipeline based on Hexagonal Architecture to solve the trade-off between accuracy and latency. The system combines a synchronous rule-based engine (Tier 1) for immediate violation processing and an asynchronous LLM-based pipeline (Tier 2) for complex contextual analysis, with Kafka Streams-based batch processing to optimize cost and throughput. The proposed pipeline achieved a Macro-F1 score of 0.8945 on an evaluation dataset of N=1000, with the rule-based engine showing an internal latency of 8ms (p50). Acknowledgment: Language editing was assisted by AI-based tools (ChatGPT, Claude).
Im Sin Cheong (Mon,) studied this question.