Harnessing Deep Learning to Detect Irony and Sarcasm in News Headlines for Combating Misinformation in Digital Media

Digital news environments are increasingly shaped by algorithmic amplification and fragmented audience engagement, enabling the unchecked spread of misinformation. Among the rhetorical strategies that obscure truth, irony and sarcasm pose unique challenges to automated detection systems due to their subtle contextual dependencies and linguistic ambiguity. To better understand and mitigate these forms of obfuscation, we curate a dataset of 1,200,000 English-language news headlines, combining verified satirical sources and crowd-sourced annotations to capture latent sarcastic and ironic cues. Among several transformer models evaluated, XLNet achieved the strongest performance and forms the basis of the primary reported results. We further apply a dual-layered attention mechanism to differentiate between ironic critique and factual distortion. To examine affective undertones, we incorporate the VADER sentiment lexicon to profile emotional valence and juxtapose it with misclassification likelihoods. Results reveal that sarcasm often overlaps with misinformation indicators, particularly when humor is weaponized to delegitimize opposing viewpoints or obscure factual content. We identify four recurring sarcasm archetypes in misinformation-laden headlines, with varying degrees of recognizability to models and readers alike. Contrary to expectations, emotional polarity alone proved insufficient for accurate sarcasm detection, suggesting a necessary interplay between pragmatics and machine inference. Our findings highlight the role of nuanced language in complicating efforts to regulate misinformation and offer empirical insight for developers of trustworthy news-ranking algorithms and digital literacy tools.