Comparative Analysis of Large Language Model in Assisting Cross-domain Rumor Detection

Rumors have certain harmfulness and variability, but cross-domain rumor detection faces challenges such as the scarcity of target domain annotation and the sensitivity of domain offset. Large language model (LLM) provides a new solution for this task with its extensive general knowledge and semantic understanding ability. This paper systematically reviews three cross-domain rumor detection methods driven by LLM, namely, the MONTROSE method based on data synthesis, the T²ARD method based on pseudo-label annotation, and the RAEmoLLM method based on retrieval enhancement. The specific role of LLM in cross-domain applications is analyzed according to its generation ability, pseudo-label annotation ability and context reasoning ability. This paper also compares the similarities and differences of the three methods in terms of call timing, computational overhead, domain adaptation mechanism, and it discusses the challenges faced by the current methods in terms of credibility. This paper aims to provide a theoretical reference for the construction of a lighter and more reliable cross-domain rumor detection system.