Abstract:Content moderation in online platforms faces persistent challenges as a result of evolving complexity of person-generated content and the constraints of traditional rule-primarily based and machine studying approaches.
While current advances in giant language models (LLMs) have enabled more sophisticated moderation via direct prompting or high quality-tuning, these approaches often exhibit limited generalization, interpretability, and adaptableness to unseen or ambiguous instances.
In this work, we suggest a novel moderation framework that leverages analogical examples to reinforce rule induction and choice reliability. Our method integrates end-to-end optimization of analogical retrieval, rule technology, and moderation classification, enabling the dynamic adaptation of moderation rules to various content situations.
Through comprehensive experiments, we show that our method significantly outperforms each rule-injected high-quality-tuning baselines and multi-stage static RAG pipelines in terms of moderation accuracy and rule high quality. Further evaluations, together with human assessments and external mannequin generalization checks, verify that our framework produces rules with better clarity, interpretability, and applicability.
These findings present that analogical instance-pushed methods can advance strong, explainable, and generalizable content material moderation in actual-world applications.