This paper presents a novel framework that leverages annotation inconsistencies in content moderation. Existing content moderation systems combine human moderators with machine learning models, but tend to treat annotation inconsistencies as noise. This paper interprets these inconsistencies as valuable signals revealing content ambiguity and presents an approach that simultaneously learns toxicity classification and annotation inconsistencies through multi-task learning. Specifically, it leverages conformal prediction to account for annotation ambiguity and model uncertainty, providing moderators with the flexibility to adjust thresholds for annotation inconsistencies. Experimental results show that the proposed framework improves model performance, calibration, and uncertainty estimation compared to single-task approaches, increases parameter efficiency, and enhances the review process.
