Clickbait Detection via Prompt-Tuning With Titles Only

Clickbait refers to deliberately enticing users into clicking through sensational text, and the corresponding content is usually totally unrelated. Recently, clickbait plays a semantic deceptive role on online web applications like news portals and social media, which will inevitably lead to user dissatisfaction and agitation. The recent clickbait detection approaches typically require large-scaled labeled data or auxiliary information like news content, however, humans can distinguish whether they are clickbaits without seeing the labeled content but only based on the titles. In this paper, we present a Clickbait Detection method that makes use of Prompt-tuning which can only leverage the few-shot labeled titles to train the detection model. Specifically, we feed the titles into the prompt template to predict the masked label, and we conducted five strategies to expand label words space for final detection. Despite being few-shot labeled titles, experimental results show that our method can obtain an obvious improvement compared to other clickbait detection methods based on deep neural networks and pre-trained language models, outperforming the state-of-the-art and up to 95% in all four metrics on three well-known benchmarks.