【摘要】 [目的/意义]社交媒体平台上充斥着大量虚假信息、恶意链接及仇恨言论等有害内容，严重损害了用户体验。为了从根本上遏制有害内容的传播，识别并遏制发布此类信息的仇恨言论群体至关重要。[方法/过程]从用户属性、用户生成内容、社交网络关系构建仇恨言论群体特征体系。用户属性包括性别和账号信息，用户生成内容涉及历史转发频率、情感和主题分析，社交网络采集交互特征和高影响力用户主题特征。进而引入基于二元组约束的相似度学习法，借助全连接神经网络与对比损失函数，构建仇恨言论群体发现模型。[结果/结论 ]仇恨言论群体特征体系具有一定的普适性，以该特征体系为基础构建的基于相似度学习的社交媒体仇恨言论群体发现模型预测准确率达到97.38%。相较于其他的分类模型，基于相似度学习的社交媒体仇恨言论群体发现模型能够更精准地识别社交媒体中的仇恨言论群体。更多还原

【Abstract】 [Purpose/Significance] Social media platforms are flooded with a large amount of false information, malicious links, hateful speech, and other harmful content, which seriously harms the user experience. To fundamentally curb the spread of harmful content, it is crucial to identify and suppress hate speech groups that post such information. [Method/Process] This paper constructed a feature system for hate speech groups based on user attributes, content generation, and social network relationships. User attributes included gender and account information. User-generated content involved historical retweet frequency, sentiment, and topic analysis. Social networks collected interaction features and topic features of high-influence users. And it innovatively introduced a similarity learning method based on pairwise constraints. It adopted a fully connected neural network and a contrastive loss function and constructed a discovery model for hate speech groups on social media. [Result/Conclusion] The feature system for hate speech groups proposed in this paper has a certain universality, and the prediction accuracy of the similarity learning-based discovery model based on this characteristic system reaches 97.38%. Compared with other classification models, the similarity learning-based discovery model can more accurately identify hate speech groups on social media.更多还原