基于噪声检测的多语言知识图谱实体对齐技术研究

针对在实体对齐任务中，由于缺少噪音实体对的标记，导致对齐准确率不高的问题，提出采用健壮性实体对齐(Robust Entity Alignment,REA)方法，设计了噪声感知实体对齐模块和噪声检测模块.首先，噪声感知实体对齐模块是基于图卷积神经网络(Graph Convolutional Networks,GCN)的知识图编码器，将知识图谱中的实体对更新嵌入；然后，基于生成对抗网络(Generative Adversarial Networks,GAN)设计了噪声生成器和噪声鉴别器，从而将实体对中的噪音实体对区分出来；最后，通过一种交互的强化训练策略，迭代使噪声感知和实体对齐相结合.实验结果表明，在DBP15K数据集上测试，新方法能有效提高在涉及噪音情况下的实体对齐精准度，与GCN-Align和IPTransE这些基准嵌入模型相比，Hits@1、Hits@5、M_RR 3个评价指标上均有较大的提升.

Research on entity alignment technology of multilingual knowledge map based on noise detection

In the entity alignment task, the accuracy of alignment is disturbed due to the lack of labels for noisy entity pairs in the entity alignment task. Robust Entity Alignment (REA) method is proposed, and noise sensing entity alignment module and noise detection module are designed. The noise sensing entity alignment module is a knowledge map encoder based on Graph Convolutional Networks (GCN), which updates and embeds entity pairs in the knowledge map. The noise detection module designs a noise generator and a noise discriminator based on the Generic Adversary Networks (GAN) to distinguish the noise entity pairs in the entity pairs. Finally, an interactive reinforcement training strategy is used to combine iterative noise perception with entity alignment. The experimental results show that the new method can effectively improve the accuracy of entity alignment in the case of noise when tested on the DBP15K dataset. Compared with GCN Align and IPTransE benchmark embedded models Hits@1、Hits@5 and M_RR evaluation indicators have been greatly improved.