Exploration and Evaluation of Prompting Methods for Text Style Transfer

Text Style Transfer is a task concerned with modifying the attributes of a text while leaving its meaning unchanged. In recent years, this task has gained much attention due to the promising performance of deep learning models. However, progress is still slow due to an overreliance on large datasets and a lack of reliable evaluation methodology. To address the data challenge, the present research proposes the usage of prompting techniques which have shown very good performances on other natural language generation tasks in few-shot settings. Furthermore, considering the lack of conventional evaluation methods, this study explores various methods employed in previous research. This includes automatic metrics, but also human evaluation. Prompting outperforms fine-tuning according to computed metrics and to human assessment, but there seems to be little correlation between the human judgement and automatic metrics assessing then meaning conservation and the fluency of generated text.