Transformer-based conformal predictors for paraphrase detection

Transformer architectures have established themselves as the state-of-the-art in many areas of natural language processing (NLP), including paraphrase detection (PD). However, they do not include a confidence estimation for each prediction and, in many cases, the applied models are poorly calibrated. These features are essential for numerous real-world applications. For example, in those cases when PD is used for sensitive tasks, like plagiarism detection, hate speech recognition or in medical NLP, mistakes might be very costly. In this work we build several variants of transformer- based conformal predictors and study their behaviour on a standard PD dataset. We show that our models are able to produce \emph{valid} predictions while retaining the accuracy of the original transformer-based models. The proposed technique can be extended to many more NLP problems that are currently being investigated.