Neural Text Rewriting: Style Transfer, Figurative Language, and Beyond

Neural networks have yielded great breakthroughs in NLP in recent years, but the vast majority of research has focused on literal language, while modelling text attributes, or style has received less attention. In this thesis, we focus on Natural Language Generation and zoom in on controllable text rewriting, exploring how to model and manipulate the style of the input text, automatically generating a new text.

Based on specific NLP tasks, we aim to address several challenges in neural text rewriting. First, we propose various efficient learning methods, including data- and model-efficient, for style transfer across different tasks and languages, providing useful insights for low-resource scenarios. Second, we put forward useful practices for automatic evaluation in style transfer, advance the establishment of still much needed evaluation standards, and provide a first glimpse into the role of large language models (LLMs) in the multidimensional evaluation of stylized text generation. Third, we take a significant first step towards the unified modelling of non-literal language across multiple figures of speech and languages, which exploits potential knowledge transfer across figures of speech. We present two novel frameworks for figurative language detection and generation, respectively. Lastly, we explore two related text rewriting tasks beyond more classic text style transfer and figurative language: (i) we introduce the novel task of responsibility perspective transfer, investigating the potential of using LLMs for raising awareness of perspective-based writing and reformulating text to that end; (ii) we propose a novel language-meaning modelling framework for meaning-to-text generation in a multilingual context.