High Disagreement

High disagreement in human annotations, particularly in subjective tasks like legal case classification and sentiment analysis of social media, presents a significant challenge for machine learning. Current research focuses on developing models that can effectively learn from datasets containing substantial human disagreement, exploring techniques like incorporating multiple ground truths and leveraging graph neural networks to capture complex interaction patterns within data. These efforts aim to improve model accuracy, calibration, and explainability, ultimately leading to more robust and trustworthy AI systems for applications ranging from legal tech to social media moderation.