Federated Learning for Exploiting Annotators’ Disagreements in Natural Language Processing