Kakao Introduces Auxiliary LLM for Text Monitoring to Enhance Efficiency by Explaining Operator Decision Basis

Kakao has introduced an LLM for text-based harmful content monitoring through an internal project called 'jamon'. It is designed as an auxiliary tool to detect hidden illegal promotional phrases in long posts and assist operators in making decisions.

Operator-Centric Assistant Role

Moving away from initial automatic blocking, the system emphasizes explanation features to help operators understand the 'reason' why content is deemed harmful. Instead of simple classification, the LLM provides regulatory status, summary, keywords, and regulatory reasons in JSON format, reducing unnecessary re-review time caused by false positives and improving work efficiency.

Model Architecture and Training Strategy

Based on a model with 8 billion parameters, it was optimized for stable inference in an A100 GPU environment. A two-track strategy was applied for real-time classification (0/1 response) and monitoring assistance (detailed JSON explanation). The effectiveness of data augmentation for classification boundaries was confirmed through strategies for expression diversification and resolving data imbalance.

Performance and Impact

Providing structured analysis results reduced the time spent deliberating during monitoring by more than 20% compared to before. Additionally, Recall for spam variations significantly improved, enabling better detection of various spam patterns, and contributed to reducing operator fatigue caused by false positives in actual operational environments.