为实现中文网络安全威胁情报知识图谱的构建，探索一种融合改进的Focal Loss和多粒度卷积神经网络的多分类方法，对情报中文本的实体关系进行抽取。针对中文网络安全威胁情报多种关系类型中存在的长尾问题，通过改进Focal Loss损失函数，增强对难区分样本的学习，提高模型的分类能力；针对情报内中英文词汇混杂以及专业词汇众多导致的关键词汇长度的方差较大，模型学习困难的问题，提出采用多粒度卷积神经网络的方法捕捉不同粒度的语句特征，提升模型在分类任务上的效果。对比实验表明，与在其他领域常用的分类方法相比，所提出的MCNNFL模型的Weighted-F1值和正确率显著提高，提升了中文网络安全威胁情报文本实体关系抽取的效果。消融实验表明，针对上述两个问题提出的两种方法皆可提升模型性能，且可以同时使用。

To achieve the construction of a Chinese cyber threat intelligence knowledge graph, a multi-classification approach was explored that combined an enhanced focal loss with a multi-scale convolutional neural network. This approach was used to extract entity relationships from the text. To address the long-tail problem within Chinese cyber threat intelligence relation extraction, the Focal loss function was improved to enhance the learning of challenging samples. To tackle the problem of significant variance in keyword lengths caused by the mixture of Chinese and English vocabulary and the abundance of specialized terms in intelligence, a multi-granularity convolutional neural network was employed to capture features of statements at different scales. This approach enhanced the model's performance in the task. The comparative experiments demonstrate that the Weighted-F1 score and accuracy of the proposed MCNNFL model are significantly higher compared to the methods commonly used in other fields. Ablation experiments indicate that both proposed methods for the problems mentioned above can improve model performance and boost performance simultaneously.