Is it possible to update the weights of a vanilla transformer model using counterexamples alongside examples?
For example, from the PAWS data set, given the phrases "Although interchangeable, the body pieces on the 2 cars are not similar." and "Although similar, the body parts are not interchangeable on the 2 cars." we have the label 0 because it is a counterexample, whereas for the phrases "Katz was born in Sweden in 1947 and moved to New York City at the age of 1." and "Katz was born in 1947 in Sweden and moved to New York at the age of one." we have the label 1 because it is a positive example of a valid paraphrase.
My goal is to use the transformer model to generate paraphrases, and I am attempting to build a GAN but could not find any references for updating the transformer text-to-text model using counterexamples.