Exploring the Reformulation of NLP Tasks as Text Generation Tasks

Abstract

In recent years, NLP classification tasks have been reformulated as text generation tasks in the form of text-to-text transformer-based models that achieve state-of-the-art performance by better utilizing pre-trained language models. This work provides a historical background, a taxonomy based on the output structures of these methods, an exploration of aspects of such models with several representative works, and discusses the current state and future of these models.