With a proper set of hypotheses labels describing the labels of the classification task, the model can infer by semantic similarity how to categorize a text sequence without having been explicitly trained for that task. For example, a sentiment analysis task could be formatted as follows:
...