Unsupervised Style Transfer in News Headlines via Discrete Style Space

Unsupervised Style Transfer in News Headlines via Discrete Style Space Qianhui Liu, Yang Gao∗, Yizhe Yang School of Computer Science and Technology, Beijing Institute of Technology, Beijing, China {3120201048,gyang,yizheyang}@bit.edu.cn Abstract The goal of headline style transfer in this paper is to make a headline more attractive while maintaining its meaning. The absence of parallel training data is one of the main problems in this field. In this work, we design a discrete style space for unsupervised headline style transfer, short for D-HST. This model decomposes the style-dependent text generation into content-feature extraction and style modelling. Then, generation decoder receives input from content, style, and their mixing components. In particular, it is considered that textual style signal is more abstract than the text itself. Therefore, we propose to model the style representation space as a discrete space, and each discrete point corresponds to a particular category of the styles that can be elicited by syntactic structure. Finally, we provide a new style-transfer dataset, named as TechST, which focuses on transferring news headline into those that are more eye-catching in technical social media. In the experiments, we develop two automatic evaluation metrics — style transfer rate (STR) and style-content trade-off (SCT) — along with a few traditional criteria to assess the overall effectiveness of the style transfer. In addition, the human evaluation is thoroughly conducted in terms of assessing the generation quality and creatively mimicking a scenario in which a user clicks on appealing headlines to determine the click-through rate. Our results indicate the D-HST achieves state-of-the-art results in these comprehensive evaluations. 1 Introduction A style makes sense under pragmatic use and becomes a protocol to regularize the manner of commu- nication [Jin et al.2022, Khalid and Srinivasan2020]. So, the task of text style transfer is to paraphrase the source text in a desired style-relevant application [Toshevska and Gievska2021]. In practical use, the style is data-driven and task-oriented in different area [Jin et al.2022]. The absence of parallel training data for a certain style is one of the difficult problems. Continuous latent space mapping is a typical method for unsupervised style transfer to address the issue. Guo et al. (2021; Liu et al. (2020) model the latent space to a Gaussian distribution. Points in latent space are moved to the target representation with some style guidance. Nangi et al. (2021; John et al. (2018; Romanov et al. (2018) disentangle the continuous latent representation purely according to its content, and replace the source attribute to the target one. However, there are two problems of the continuous space approach. Firstly, the style is highly abstract so that it is unstable and too sparse to accurately represent the style in the continuous space. Second, the continuous vector-based representation is difficult to manipulate and cannot be examined at a finer level. To control the style transfer and enhance its explainability, several kinds of discrete signals are used to represent the style. For instance, Reid and Zhong (2021; Tran et al. (2020; Li et al. (2018) employ …

Content truncated to 500 words