Most existing entity alignment solutions rely on clean labeled data to train models, with little attention to seed noise. To address the noise problem in entity alignment, this paper proposes a labeling noise reduction framework that injects auxiliary knowledge and incidental supervision in entity alignment to correct the seed errors in the labeling and bootstrapping process. In particular, considering the weaknesses of previous neighborhood-based embedding approaches, this paper applies a new dual relational at tent ion-matching encoder to accelerate the structure learning of the knowledge graph while using auxiliary knowledge to compensate for the lack of structural representations. Then, weakly supervised label noise reduction is performed by adversarial training. For the problem of error accumulation, this paper further uses the label refinement module to improve the performance of the model. Experimental results show that the proposed framework can easily cope with the entity alignment problem in noise-laden environments and consistently outperforms other baseline methods in terms of alignment accuracy and noise discrimination on multiple real datasets.
关键词
图对齐，标签精化,降噪算法，多模态监督
Key words
图对齐，标签精化,降噪算法，多模态监督
引用本文复制引用
会议名称
Chinese national conference on computational linguistic
会议地点
Nanchang(CN)
会议母体文献
The 21st Chinese national conference on computational linguistic
NETL