How Positive Are You: Text Style Transfer using Adaptive Style Embedding

Proceedings of the 28th International Conference on Computational Linguistics, pages 2115–2125 Barcelona, Spain (Online), December 8-13, 2020 2115 How Positive Are You: Text Style Transfer using Adaptive Style Embedding Heejin Kim1 and Kyung-Ah Sohn1,2∗ 1Department of Artificial Intelligence 2Department of Software and Computer Engineering Ajou University {heeeee123, kasohn}@ajou.ac.kr ∗ Corresponding author Abstract The prevalent approach for unsupervised text style transfer is disentanglement between content and style. However, it is difficult to completely separate style information from the content. Other approaches allow the latent text representation to contain style and the target style to affect the generated output more than the latent representation does. In both approaches, however, it is impossible to adjust the strength of the style in the generated output. Moreover, those previ- ous approaches typically perform both the sentence reconstruction and style control tasks in a single model, which complicates the overall architecture. In this paper, we address these issues by separating the model into a sentence reconstruction module and a style module. We use the Transformer-based autoencoder model for sentence reconstruction and the adaptive style embed- ding is learned directly in the style module. Because of this separation, each module can better focus on its own task. Moreover, we can vary the style strength of the generated sentence by changing the style of the embedding expression. Therefore, our approach not only controls the strength of the style, but also simplifies the model architecture. Experimental results show that our approach achieves better style transfer performance and content preservation than previous approaches.1 1 Introduction Text style transfer is the task of modifying a text with a specified style attribute. Given a source text with an attribute (e.g., positive), the text style transfer problem aims to alter this input text to achieve a different attribute (e.g., negative). The significant concern in this problem is that the generated text must meet some requirements: the text must (i) reflect the given target attribute, (ii) maintain the content that is style-irrelevant, and (iii) generate a sentence that appears natural. Text style transfer has been solved by a supervised method using a parallel dataset (Jhamtani et al., 2017) containing pairs of source and target sentences. However, obtaining this parallel dataset that achieves a total one-to-one correspondence with a specified style is often not possible. Therefore, un- supervised approaches have recently been actively researched. The first among these unsupervised ap- proaches is disentanglement. To satisfy the first two style transfer requirements, namely, reflecting the style attributes and preserving the contents, the disentanglement approach tries to perfectly separate the style component from the content component within the input text. If the input is completely separated, the text style is transferred by substituting the style component with a new target; for this approach, a discriminator is used to separate the style from a latent representation z, a compressed representation of the input text (Fu et al., 2018; John et al., 2019). The discriminator can also be applied to the output text so that the output text reflects …

Content truncated to 500 words