Text Style Transfer Based On Combined Style Representation And Cycle Reconstruction With Content Alignment

As a controllable text generation task,text style transfer is easy to be integrated with other natural language processing task,which is defined to transform text from one style to another while keeping the semantic content of the text unchanged,and it has important research value.We need to use unsupervised learning method for text style transfer because the main difficulty is the lack of parallel texts with different styles and the same content.This thesis focuses on the unsupervised text style transfer based on combined style representation and cycle reconstruction with content alignment.An unsupervised text style transfer model based on the variational autoencoder is designed.Through the adversarial training between the discriminator and variational autoencoder,the content and style of the source sentence are separated in the latent space,and the style-independent text content representation is obtained.The target style representation and the style-independent text content representation output by the encoder are used as the inputs of the generator to generate the target style sentence.In this thesis,a style combined representation method is proposed to address the shortcoming of the method where a linear transformation of a binary vector is used as the representation of the style,which limits the representation of the text style and increases the burden on the generator.Two encoders are designed to extract the content and style representation of the text in the latent space.The latent style variable from the style encoder and the style vector obtained by linear transformation of the style label are combined as the final style representation.On the datasets Yelp and GYAFC,the correctness of the combined style representation method proposed in this thesis has been verified,and compared to the CA and CG models,the style combined representation method has a certain improvement in the accuracy of text style transfer and the text fluency.To address the problem of low content preservation quality of style combined representation method,a cycle reconstruction based on content alignment method is proposed in this thesis.Transfer a text to another style and then transfer it back to original style.In this process,the cycle reconstruction loss is constructed by constraining the consistency between the original text and the text through twice style transfer,and the content alignment loss is constructed by constraining the consistency of the content implicit variables at the same time,so as to improve the content preservation quality of text style transfer.The experimental results on two datasets Yelp and GYAFC show that the degree of text content preservation has been improved by the method of the cycle reconstruction with content alignment.