推荐系统在解决新闻准确呈现的问题上显示出巨大的潜力。现有的新闻推荐系统大多只考虑新闻文本，忽略了新闻图片与用户之间的关系。然而，新闻图片也是用户决定点击新闻的重要因素。本文将ViLBERT与多模态知识图注意力网络相结合，利用多模态知识提高新闻推荐系统的准确率，使用多模态图关注技术在多模态知识图关注网络上传播信息，将生成的图像和文本聚合嵌入推荐的表示，以有效地表征目标，缓解推荐系统中用户行为稀疏和冷启动的问题。通过在两个不同的真实中英文新闻数据集上进行了实验，结果表明本模型可以有效地提高新闻推荐。

Recommender systems have shown great potential to solve the problem of accurate presentation of news. Most of the existing news recommender systems only consider the news texts but ignore the relationship between the news picture and user. However, news images are also a significant factor in users' decision to click on news. In this paper, we proposed ViLBERT combined with Multi-modal Knowledge Graphs Attention Network (VMKGAT) to better enhance the accuracy of the news recommender system by using multi-modal knowledge. We used a multi-modal graph attention technique to disseminate information on the multi-modal knowledge graph attention network, and then used the generated images and text aggregation to embed the representation for the recommendation, It could effectively characterize the target item and alleviate the problems of sparse user behavior and cold start in recommendation system. We conducted a large number of experiments on two different real English and Chinese news datasets, and the experimental results show that our model VMKGAT can effectively improve news recommendation.