1) The document presents an approach called Multidimensional Annotation Scaling (MAS) for aggregating complex annotations from multiple annotators. 2) MAS models annotation tasks as distance matrices calculated using task-specific distance functions, rather than modeling the annotations directly. 3) It then applies a Bayesian hierarchical model called multidimensional scaling to learn annotator reliabilities and item difficulties from the distance matrices in order to aggregate the annotations. 4) Experiments on tasks with diverse complex label types like sequences, rankings and translations show MAS outperforms baselines and adapts to different tasks without retraining.