Cross-model diffusion: Mitigating hallucination in large language models for rumor detection.

Abstract

Various emergencies in the current digital age are frequently accompanied by the rapid spread of numerous rumors, making accurate identification crucial to reducing public panic and misinformation. Existing research methods mainly rely on graph neural networks to model the propagation structure. The understanding of deep semantic information in text by these methods is limited, making it difficult to learn high-quality representations. To obtain high-quality representations required for rumor detection, we propose the cross-model diffusion framework. The framework fuses representations from large language models and pre-trained language models through a diffusion process to achieve complementary enhancement. Specifically, we leverage the strong comprehension and reasoning capabilities of large language models to extract deep semantic information from text. Meanwhile, to mitigate the hallucination effects of large language models, we introduce pre-trained language models, leveraging their advantage in textual consistency as a complement. To facilitate effective semantic collaboration between models, we propose a diffusion-based fusion approach that enables deep semantic complementarity by coordinating inter-model representational disparities. Experimental results on cross-lingual benchmark datasets verify the effectiveness of the proposed framework.