Few-shot rumor detection based on prompt learning and meta learning

The rapid development of Internet technology makes online rumors increasingly rampant. The current researches on rumor detection rely on large-scale labeled instance data. However, in the face of emergent events, obtaining large-scale labeled data becomes extremely difficult. To address this issue, this paper proposed a few-shot rumor detection model based on prompt learning and meta learning, named MPRD (Mate Prompt Rumor Detection, MPRD). MPRD combines hard prompt and soft prompt to explore universal rumor prompt templates that are not limited by specific topics, thus offering greater flexibility. Moreover, it employs meta-learning methods to sample meta-tasks based on past events, trains prompt models on these meta-tasks, and ultimately applies them to perform rumor detection tasks in emergency events. Experiments on two publicly datasets show that MPRD outperforms state-of-the-art baseline models on detection accuracy, and exhibits significant performance improvement on few-shot rumor detection under different topic.