Text Style Transfer Using DRG Framework of Combined Retrieval Strategy

Published: 01 Jan 2021, Last Modified: 05 Feb 2025BigComp 2021EveryoneRevisionsBibTeXCC BY-SA 4.0
Abstract: Text style transfer is a task that changes the stylistic attribute while preserving content information. It can be used for various applications such as intelligent chatbot and literary works. Based on the recently appeared model, namely General Style Transformer(GST), we propose a new variant of it by replacing one of its step. The GST basically follows three steps: delete, retrieve, and generative (DRG) steps. The DRG framework is a pipeline structure, so errors of the deletion step might propagate to its following steps. To mitigate this, we modify the retrieval step to take the result sentence of the deletion step and the raw sentence as well, and results of the two sentences. It combines the results of the two sentences, so it has a chance to alleviate some errors generated by the deletion step. By experimental results with Yelp dataset, we prove the effectiveness of our approach.
Loading