Exploiting Paraphrasers and Inverse Paraphrasers: A Novel Approach to Enhance English Writing Fluency through Improved Style Transfer Training Data
Zhendong Du, Kenji Hashimoto · 2023
In the realm of enhancing English writing fluency, the scarcity of high-quality training data has perennially posed a significant challenge. Moreover, elevating the fluency of writing while ensuring the preservation of semantic integrity compounds the intricacies of this task. In this study, we introduce and implement a style converter rooted in the Paraphraser and Inverse Paraphraser methodologies, aimed at ameliorating English writing fluency. Concurrently, this converter facilitated the generation of a voluminous corpus of synthetic training data. Utilizing this data, we fine-tuned GPT-2 to forge an English text style transfer model. Remarkably, despite our model being trained on a dataset substantially smaller than that of prevailing baseline methods, it exhibited exemplary performance across multiple evaluation metrics, even surpassing these baselines on certain pivotal indices. These findings corroborate the efficacy of our approach and underscore its immense potential in the domain of English writing fluency enhancement. This investigation not only offers a novel optimization strategy for English composition but also furnishes researchers in cognate fields with fresh research perspectives and methodologies.