Style Transfer Based on Discrete Space Modeling

This paper is dedicated to exploring headline style transfer techniques, with the core objective of enhancing the appeal of headlines while preserving their semantics. A significant challenge is the lack of parallel training datasets. To address this issue, we propose an unsupervised headline style transfer model called the Discrete Headline Generation (DHG) model. This model decomposes the process of generating text with specific styles into two key steps: content feature extraction and style modeling. Two text encoders separately extract the content information and style information from the input text, while a text decoder integrates the knowledge from both encoders to generate content with a specific style.Given that style signals are more abstract compared to textual content, we propose constructing the style representation space as a discrete space, where each discrete point corresponds to a specific style category. Experimental results demonstrate that the DHG model exhibits state-of-the-art performance in comprehensive evaluations.