In zero-shot learning, a learner (trained on some other data) needs to predict the class of a sample without having observed examples of the class during training. Naturally, the learner / model cannot generate classes without any prior information whatsoever, so it needs to be provided general information about these zero-shot classes.