Is there Binary Zero-Shot Learning with no defined prototypes for the unseen class?

I deal with a text classification problem, where there are only two classes; relevant and irrelevant. That is, a text might be relevant / irrelevant with a predefined topic. I have a dataset that consists of only relevant text instances. I am thinking of using zero-shot learning. A way to represent the semantic space (prototype) for the relevant instances is to use word embeddings and to encode the text in a fixed size vector. But how can I represent the semantic space (prototype) for the irrelevant class which should represent all instances that are not relevant ?