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

Area of Science:

• Computational Linguistics
• Artificial Intelligence
• Information Science

Background:

• The rapid spread of digital-age rumors necessitates accurate identification to combat public panic and misinformation.
• Current graph neural network methods for rumor detection have limited understanding of deep semantic text information, hindering high-quality representation learning.

Purpose of the Study:

• To propose a novel cross-model diffusion framework for enhancing rumor detection.
• To fuse representations from large language models (LLMs) and pre-trained language models (PLMs) for complementary representation enhancement.
• To achieve high-quality text representations crucial for accurate rumor detection.

Main Methods:

• Leveraging LLMs for deep semantic understanding and reasoning capabilities.
• Incorporating PLMs to mitigate LLM hallucination effects and ensure textual consistency.
• Employing a diffusion-based fusion approach to coordinate inter-model representational disparities and enable semantic collaboration.

Main Results:

• The proposed framework effectively fuses representations from LLMs and PLMs.
• Diffusion-based fusion facilitates deep semantic complementarity between models.
• Experimental results on cross-lingual benchmark datasets demonstrate the framework's effectiveness in rumor detection.

Conclusions:

• The cross-model diffusion framework significantly improves the quality of text representations for rumor detection.
• Combining the strengths of LLMs and PLMs through diffusion offers a promising approach to mitigate misinformation.
• The method shows effectiveness in cross-lingual rumor detection scenarios.