GTAE: Graph Transformer–Based Auto-Encoders for Linguistic-Constrained Text Style Transfer

Non-parallel text style transfer has attracted increasing research interests in recent years. Despite successes transferring the based on encoder-decoder framework, current approaches still lack ability to preserve content and even logic of original sentences, mainly due large unconstrained model space or too simplified assumptions latent embedding space. Since language itself is an intelligent product humans with certain grammars a limited rule-based by its nature, relieving this problem requires reconciling capacity deep neural networks intrinsic constraints from human linguistic rules. To end, we propose method called Graph Transformer–based Auto-Encoder, which models sentence as graph performs feature extraction at level, maximally retain structure sentences. Quantitative experiment results three non-parallel tasks show that our outperforms state-of-the-art methods preservation, while achieving comparable performance accuracy naturalness.