A recent line of research in NLP proposes to never aggregate human annotations 1 , but rather to leverage the worth of knowledge found in label variation for building models [1] and evaluating them [2]. This approach is particularly relevant when dealing with highly subjective aspects of natural language such as irony or undesirable language. In this talk, I will present the perspectivist paradigm in NLP, and the results of recent and ongoing research focusing on building perspective-aware predictive models and automatically extract human perspectives from annotated data. Particular emphasis will be given to language resources, for training or few-shotting models, but also for benchmarking in the perspectivist framework.