STUDYING FORMALITY STYLE TRANSFER USING SEQ2SEQ MODEL AND PRE-TRAINED NEURAL NETWORKS

This research study is an attempt to effectively and accurately transfer the formality of a sentence and discusses the combination of a pre-trained GPT-2 language model with a rule based system using a single self-attention mechanism. The dataset used is Grammarly's Yahoo Answers Formality Corpus (GYAFC), 2018. The corpus consists of informal and corresponding formal sentences in parallel and the model discussed generates a 93% accuracy on the GYAFC testing data.