Text Style Transfer in Italian Newspaper Articles

Recent advances in deep generative models have yielded remarkable success in various Natural Language Processing (NLP) tasks. Text Style Transfer (TST) is one such task, involving the modification of text to control its stylistic features while preserving the underlying semantic meaning. The most frequent instances of the task target attributes such as sentiment, formality, politeness, and authorship, among others. This thesis presents a detailed review of the current state-of-the-art methods for TST and applies some of those methods to the context of processing Italian news articles. Specifically, the aim is to update the style of news articles written by certain newspapers to mimic the writing style of others. To achieve this, various techniques are implemented and compared on a non-parallel corpus of news articles from multiple sources, with evaluations conducted using automated methods proposed in the literature. The ultimate objective is to create a compelling and robust style transfer pipeline that can be applied in a real-world scenario, laying the groundwork for future work in this field.