Abstract: To address the limitations of existing methods for detecting internet memes because of their irregular expressions, rapid propagation, and ambiguous semantics, this study proposes a collaborative framework that integrates a dynamic knowledge graph with retrieval-augmented generation. By constructing an adaptively updated knowledge graph and leveraging the contextual understanding capabilities of large language models, the proposed framework enables accurate identification and evolutionary trajectory tracking of internet memes. Experiments conducted on a self-built social media corpus demonstrate that the proposed method effectively overcomes the time-lag limitations of traditional static models and the lack of interpretability inherent in black-box models. It achieves superior performance in memes detection compared to mainstream baseline models. Furthermore, the study provides an in-depth analysis of the advantages and challenges of the graph-based retrieval-augmented strategy in handling noisy and unconventional expressions, and systematically validates the contribution of the proposed framework to overall performance. Finally, a case study of real-world internet memes evolution is presented to verify the effectiveness of the framework, offering support for real-time perception and in-depth analysis of highly noisy and rapidly evolving internet language phenomena.

Key words: Dynamic knowledge graph, Retrieval-augmented generation, Large language models, Internet Meme, Semantic association analysis