Using BART for zero-shot learning
In the field of machine learning, zero-shot learning is referred to as a model that can perform a task without explicitly being trained on it. In the case of NLP, it’s assumed that there’s a model that can predict the probability of some text being assigned to classes that are given to the model. However, the interesting part about this type of learning is that the model is not trained on these classes.
With the rise of many advanced language models that can perform transfer learning, zero-shot learning came to life. In the case of NLP, this kind of learning is performed by NLP models at test time, where the model sees samples belonging to new classes where no samples of them were seen before.
This kind of learning is usually used for classification tasks, where both the classes and the text are represented and the semantic similarity of both is compared. The represented form of these two is an embedding vector, while the similarity...
