摘要
With the explosive growth of online information, misinformation spreads rapidly and becomes increasingly difficult to detect. This creates a critical concern, as the public faces increasing exposure to fabricated content. Developing automated detection methods is essential to mitigate its influence. To address this challenge, we propose a novel framework, expert discussion network (ExpertDN), which leverages a large language model within a multiagent system featuring a structured multiround cross-debate mechanism. ExpertDN instantiates domain-specific expert agents that simulate collaborative reasoning through iterative critique and self-updating. These interactions are structured into a heterogeneous graph, capturing both semantic content and the reasoning dynamics among agents. The resulting graph is encoded using a heterogeneous graph transformer, and the final representation is fused with textual embeddings through cross-attention and then passed to a multilayer perceptron classifier to assess the veracity of the article. We evaluate ExpertDN on four widely used misinformation datasets. Experimental results demonstrate that our proposed ExpertDN consistently outperforms existing baselines across all benchmarks.