Vérification informatique des faits et détection de la désinformation

In the era of social networks, the propagation of fake news is increasingly frequent. In the last decade, numerous events were affected by misinformation. Whether we think of the last US election, or the COVID crisis, there is little doubt about the influence of such dishonest communication. Therefore, organisms do their best to counter them. Their efforts aim at reducing fake news amount by fact-checking claims published on social medias. Traditional methods rely on humans to manually annotate each claim, for example in a politician's speech. However, this method is not adaptable to the vast amounts of data found on social media, nor to crisis times when there can be sudden increases in social media posts. Therefore, a preferred solution is computational fact checking. In this domain, the best solutions are Machine Learning based. Usually, given a claim, a model retrieves relevant evidence from a corpus. Then, a predictor model assesses the veracity of the claim based on the evidence. In this thesis, we investigate how to support information mitigators, such as fact-checkers, in their work. We focus on two main aspects that are clear bottlenecks for the adoption of computational solutions in practice. The first aspect is the manual effort to bootstrap and refine the systems. The second aspect is to make systems more interpretable. Bootstrapping a fact-checking system, with a supervised learning approach, requires a lot of training data. As manual annotation is expensive and slow, it is not adaptable to new domains, neither to crisis situations. In such context, governments or companies cannot afford to wait until a new training set is manually defined, as this may be too long. The solution we propose is to generate such datasets. Synthetic datasets are increasingly popular for model fine-tuning or training. In fact-checking, little research has been done on synthetic datasets construction from tabular evidence. In our work, TENET, we propose to handle this issue by generating claims from structured evidence sources. In UNOWN, we switch to the multi-modal setting to perform claim generation using both the tabular and the textual modalities. We provide generated examples as training data to fact-checking systems. The systems trained with our data showcase a label-prediction accuracy comparable to the same system trained with human-created training examples. Another matter with computational fact-checking is explainability. Fact-checking systems are explainable when users can understand their decisions thanks to a justification. From a claim, state of the art models predicts a label without any justification on how they reached their conclusion. Users, in a misinformation context, can be eager to know why the text they are reading is labelled as ‘refutes'. In this scenario, current black-box models fail short. In our work, we leverage two state of the art xAI approaches to interpret decision taken by models. We show that xAI techniques permit to identify relevant evidence with high accuracy. This thesis thoroughly explores Synthetic Dataset generation and Explainability of Fact-checking systems' topics. The systems we propose ease the building of supervised models in low resource domains and ease the understanding of predictions by the users.