Style versus Content: A distinction without a (learnable) difference?

Proceedings of the 28th International Conference on Computational Linguistics, pages 2169–2180 Barcelona, Spain (Online), December 8-13, 2020 2169 Style versus Content: A distinction without a (learnable) difference? Somayeh Jafaritazehjani Univ Rennes, CNRS, IRISA TU Dublin, ADAPT Centre somayeh.jafaritazehjani@irisa.fr Gw´enol´e Lecorv´e Univ Rennes, CNRS, IRISA gwenole.lecorve@irisa.fr Damien Lolive Univ Rennes, CNRS, IRISA damien.lolive@irisa.fr John D. Kelleher TU Dublin, ADAPT Centre john.d.kelleher@tudublin.ie Abstract Textual style transfer involves modifying the style of a text while preserving its content. This assumes that it is possible to separate style from content. This paper investigates whether this separation is possible. We use sentiment transfer as our case study for style transfer analysis. Our experimental methodology frames style transfer as a multi-objective problem, balancing style shift with content preservation and fluency. Due to the lack of parallel data for style transfer we employ a variety of adversarial encoder-decoder networks in our experiments. Also, we use a probing methodology to analyse how these models encode style-related features in their latent spaces. The results of our experiments which are further confirmed by a human evaluation reveal an inherent trade-off between the multiple style transfer objectives and indicate that style cannot be usefully separated from content within these style-transfer systems. 1 Introduction Style transfer can be defined as a Natural Language Generation (NLG) task where an input word se- quence (a text, a sentence, an utterance. . .) from an original style is rephrased in order to fit a target style while keeping its semantic information, leading to a style-transferred sequence. This task can be applied to many facets of the natural language: sentiment, complexity, authorship, formality, and so on. The primary use-case of textual style transfer is to improve language understanding between humans or between humans and machines by adapting messages to the social context or abilities of an interlocutor. However, work on style transfer can also provide insight into language more generally, for example by providing insight into the relationship between style and content. Indeed, a pre-requisite of style trans- fer is the presumed ability to separate style from content, and so research on style transfer can inform whether computational models can usefully learn this distinction. Research on style-transfer often focuses on one specific aspect of textual productions, (e.g., preserving the meaning of the original sequence) and disregards the others (e.g., how present is the target style in the generated output, and the fluency of the generated sentences). However, this paper highlights the im- portance of evaluating style transfer systems in terms of multiple objectives where fitting the target style is counter-balanced by the needs to generate a linguistically fluent output sequence and at the same time preserving the original meaning. Furthermore, based on an analysis of a number of evaluation experi- ments of style-transfer systems, including a probing experiment that analyses the latent space of these systems, we identify an inherent trade-off in style-transfer between transfer strength, fluency and content preservation. We conclude that style and content cannot be usefully disentangled by these systems (at least not …

Content truncated to 500 words