compositional zero-shot learning (CZSL) aims to train models to recognize
novel compositional concepts based on learned concepts such as attribute-object
combinations. One of the challenges is to model attributes interacted with
different objects, e.g., the attribute ``wet" in ``wet ap