Analysis of frames and persuasion techniques in online news articles in a multilingual setup

Persuasion is the ability to influence people's behaviour and choices by exploiting words and emotions to achieve a set goal, such as convincing the interlocutor of a message or idea. The definition has no inherently negative meaning but recognising such techniques and the context in which they are used can be essential int understanding the dynamics behind communication. Within this framework is the contribution of the mass media which, in order to quickly disseminate information or effectively send a message to as wide an audience as possible, have refined their communication strategies by using persuasion techniques, in such cases in an inflated manner. Combined with these, the skilful use of perspective framing in the description of an event and the narrative techniques that characterise a journalistic article on a narrative level allows the digital information system to be exploited for propaganda purposes. Previous works focus separately on each of these aspects. On the other hand, in this project we intend to provide a complete overview of propaganda approaches, studying all its nuances and carrying out three different analyses of multilingual texts from online newspapers. Specifically, the main aim of this work is to investigate twenty-three persuasion techniques used within news articles, accompanied by a category label and written in six different languages, to cover a wide range of topics according to fourteen possible framings. Several variants of artificial intelligence models based on the XLM-RoBERTa architecture are proposed, allowing a given article, written into one among in six languages, to be classified by category, to detect prospective framing and persuasion techniques within it. For the latter the classification is assessed at the level of the whole article and at the fine level, not only to inform about the presence of a technique but also to label it and detect the fragment of text in which it appears. Collecting textual data with a good level of information in this sense is a challenge not to be underestimated, as the supervision procedure is demanding, time-consuming and expensive in terms of resources. For this reason, Transformer-based architectures, such as mBART and ChatGPT, are used to perform data augmentation and acquire useful information to bridge language discrepancies and improve performance.