TSTG: A Text Style Transfer Model Based on Generative Adversarial Networks

Text style transfer (TST) models are gaining considerable prominence in the field of Internet of Things (IoT) applications. Nonetheless, conventional encoder-decoder frameworks encounter constraints stemming from their dependence on parallel corpora and rigid architectures. This article introduces an innovative TST model based on generative adversarial networks (TSTG) and reinforcement learning methodologies to augment the efficacy of text-style transfer. The proposed generator model utilizes a sequence to sequence (Seq2Seq) architecture to process the original text alongside the target style attributes for effective conversion, while the discriminator assesses the generated text through style judgment and ranking scores to ensure high-quality output. Moreover, a new classification model named Bidirectional encoder representations from transformers-text convolutional neural networks (BTNN) improves the style evaluation of utterances. By alternately training the generator and discriminator, this approach significantly enhances fluency and accuracy, as well as addresses common challenges including gradient propagation during the training and detection of obscure-style expressions. Comprehensive evaluations demonstrate that the model achieves superior style accuracy, content preservation, and text fluency, marking a significant advancement in TST methodologies.