Performance Comparison of Improved Common Sequence to Sequence Paraphrasing Models

Paraphrase generation is a widely known and complex task in the field of Natural Language Processing (NLP). With a variety of paraphrasing models, many researchers have developed their own methodology to overcome this task with better efficiency and quality paraphrases. In this paper, Text-to-Text Transfer Transformer (T5) and Back-Translation guided multi-round Paraphrase Generation (BTmPG), two popular algorithms in paraphrasing models, are compared. Paired paraphrased sentences are used to train both models. The goal of paraphrasing models is to generate paraphrases with a high degree of semantic similarity and the fewest possible grammatical errors and sentence structure copying. For the purpose of paraphrase generation, both the T5 and BTmPG models have been optimized. This study aims to compare the performance of the T5 and BTmPG models and identify factors that have a significant influence on output quality. The accuracy and precision of the T5 and BTmPG models' paraphrase quality are calculated using BLEU and ROUGE1. Despite requiring more time for training, the T5 model outperformed the BTmPG model slightly, according to the results. Despite having a smaller dataset size, the MRPC dataset was found to be more efficient than the QQP dataset. With proper grammar and lexicality, both models produced results that were satisfactory.