基于维基百科构建概念体系是跨语言知识图谱构建的重要研究方向之一。现有的方法将概念体系构建当作分类任务，有两个缺点：（1）依赖于英文概念体系和维基百科跨语言链接，面临错误传播问题和跨语言链接数量不足问题；（2）分类器仅仅以词条和类别的标题作为输入，难以捕获上下位语义信息。针对以上两个问题，本文提出了基于对抗训练的无监督训练方法，摆脱对英文概念体系和维基百科跨语言链接的依赖，并且提出在分类器中对维基百科类别网络的图结构特征进行编码，捕获更精准的上下位语义信息。在英文维基百科上的实验结果表明，本文提出的方法可以构建较高质量的知识图谱概念体系，效果高于当前最先进的跨语言知识图谱概念体系方法。

Taxonomy construction for cross-lingual knowledge graph based on Wikipedia is an important research area recently. The state-of-the-art method regards taxonomy construction as a binary classification problem and makes multilingual taxonomy construction possible. However, there exists two main issues: (1) existing method relies on English gold taxonomy and cross-lingual links, thus suffers from error propagation and data sparsity problem. (2) existing method cannot capture taxonomic relation accurately. We propose an unsupervised classifier based on adversarial training, thus breaking out of dependency on English gold taxonomy and cross-lingual links. The encoding of various structure features based on Wikipedia Category Network also helps to capture taxonomic relation. Experiments on English taxonomy construction demonstrate that our approach outperforms the state-of-the-art approaches.