PaPa: Propagation Pattern Enhanced Prompt Learning for Zero-Shot Rumor Detection

Heng Zhang, Peng Wu, Li Pan

Abstract: Social media platforms have significantly lowered the barriers to accessing information, which has also facilitated the proliferation and dissemination of rumors. The emergence of unforeseen breaking events on social media brings data scarcity issues to most existing rumor detection methods. To reduce dependence on large scale annotated data, prompt learning based on Pre-trained Language Models (PLMs) has emerged as a promising approach, as it bridges the gap between pre-training task of PLMs and the downstream task. Furthermore, the propagation patterns formed by social users are helpful in verifying the authenticity of emerging rumors, as they are usually hard to be fabricated and always domain-invariant. To this end, we propose PaPa, a Propagation pAttern Enhanced Prompt leArning method for zero-shot rumor detection on social media, which incorporates features from propagation patterns of social users into prompt learning framework to improve the rumor detection performance in cross-domain zero-shot scenarios. Specifically, we propose a bi-directional Time aware Graph Convolutional Network (T-GCN) to extract domain-invariant propagation structural and temporal pattern features on both forward and backward propagation directions. Then we integrate these features into a hybrid prompt learning framework by a hybrid prompt template strategy. Finally, we employ a prototypical network as the model classifier, supplanting manually labeled verbalizer. Experimental results illustrate the superiority of our approach over several baselines in zero-shot rumor detection.

External IDs:dblp:conf/iconip/ZhangWP24