Through Smoke and Mirrors of the Post-Truth Era: Knowledge-Driven Generation as Algorithmic Resistance to Misinformation

The rapid spread of misinformation online poses a growing threat to public discourse and democratic society. Research has shown that simply flagging false content is insufficient; providing users with well-grounded explanations of why a claim is false leads to more durable belief revision than mere labeling. Yet the volume at which misinformation spreads online far exceeds the capacity of manual fact-checking, making its automation not merely desirable but necessary. Natural Language Generation (NLG) represents a viable solution, enabling the automatic production of verdicts: knowledge-grounded textual responses that explain the veracity of a claim. This thesis investigates how knowledge-driven generative approaches can automate verdict generation, leveraging the growing capabilities of modern NLG technologies while accounting for the multifaceted nature of misinformation, including its diverse communication styles, linguistic and geographical spread, and the varying availability of reliable knowledge in real-world scenarios. We first cast verdict generation as a summarization task, benchmarking extractive, abstractive, and hybrid approaches and exploring multitask strategies to adapt verdict style to the guidelines of different fact-checking organizations. While effective, these systems were developed and evaluated on journalistic claims, failing to account for the social media context in which much misinformation circulates. We therefore shift to a social correction scenario, constructing a dedicated dataset and extending verdict generation to jointly adapt style and emotional register to the communication patterns of users spreading false claims. Both studies, however, remain limited to English and rely on knowledge available in the same language as the claim. We address this limitation by leveraging multilingual instruction-tuned large language models, building a professionally curated dataset across eight European languages and evaluating cross-lingual scenarios where supporting knowledge and claims are in different languages. We then turn to the more realistic setting in which no fact-checking article is directly associated with the claim, evaluating Retrieval-Augmented Generation pipelines that retrieve evidence from both homogeneous and heterogeneous knowledge bases, testing diverse retrieval strategies and examining how claim style affects pipeline performance. Finally, we recognize that not all misinformation takes the form of explicit, well-formed claims: clickbait headlines deceive through omission and require dedicated preprocessing before they can enter a fact-checking pipeline. We address this gap by proposing novel tasks and purpose-built resources for clickbait detection, spoiler generation, and neutralization. The thesis concludes by discussing the open challenges that remain in the context of counterspeech for misinformation in NLG.