Abstract: With the rapid development of social networking platforms, the issue of cyberbullying has become increasingly prominent. The diverse forms of online expression that combine text and images have increased the difficulty of detecting and managing cyberbullying. This paper constructs a Chinese multimodal cyberbullying dataset that includes both text and images. By integrating the BERT(bidirectional encoder representations from transformers) model with the ResNet50 model, we extract singlemodal features from text and images, respectively, and perform decisionlevel fusion. The fused features are then detected, achieving accurate identification of text and images as either cyberbullying or noncyberbullying. Experimental results indicate that the multimodal cyberbullying detection model proposed in this paper can effectively identify social media posts or comments that contain cyberbullying characteristics in both text and images. It enhances the practicality, accuracy, and efficiency of detecting multimodal cyberbullying, providing a new approach and method for the detection and management of cyberbullying on social networking platforms. This contributes to the creation of a healthier and more civilized online environment.

Key words: cyberbullying, multimodal, feature fusion, detection model, social network platforms

摘要： 随着社交网络平台的迅速发展，网络欺凌问题日益突出，文本与图片相结合的多样化网络表达形式提高了网络欺凌的检测和治理难度.构建了一个包含文本和图片的中文多模态网络欺凌数据集，将BERT(bidirectional encoder representations from transformers)模型与ResNet50模型相结合，分别提取文本和图片的单模态特征，并进行决策层融合，对融合后的特征进行检测，实现了对网络欺凌与非网络欺凌2个类别的文本和图片的准确识别.实验结果表明，提出的多模态网络欺凌检测模型能够有效识别出包含文本与图片的具有网络欺凌性质的社交网络帖子或者评论，提高了多模态形式网络欺凌检测的实用性、准确性和效率，为社交网络平台的网络欺凌检测和治理提供了一种新的思路和方法，有助于构建更加健康、文明的网络环境.

关键词: 网络欺凌, 多模态, 特征融合, 检测模型, 社交网络平台