Abstract: Memes exist in the form of images and texts and are used to describe hate speeches, rumors that spread among users on the Internet. They often use web entities such as popular figures, events, or historical figures to express hate emotions. These implicit emotional expressions are worth academic attention, but web entities are mostly ignored by existing meme identification methods. To address the problem, this paper proposes a meme recognition method based on graph convolutional network. Specifically, the web entity information contained in the image is first extracted. The web entity modality and the text modality are fused by a graph convolutional network. An external dictionary is employed to measure the relationship between the web entity and the meme text from multiple perspectives when building cross-domain graph. Then, the text and image modalities are interacted through the attention module. Finally, the self-distillation technology is employed to improve the model’s information utilization rate. Our experimental results on the Hateful Memes dataset and the MAMI dataset reach an accuracy of 76.03% and 73.9% respectively, and the performance is superior to the existing SOTA model.