Мультимодальное согласованное контрастивное декомпонентное обучение

AbstractWe propose to solve the natural language inference problem without any supervision from the inference labels via task-agnostic multimodal pretraining. Although recent studies of multimodal self-supervised learning also represent the …...