Improving Entity Disambiguation Using Knowledge Graph Regularization

Zhi Rui Tam, Yi-Lun Wu, Hong-Han Shuai

Abstract: Entity disambiguation plays the role on bridging between words of interest from an input text document and unique entities in a target Knowledge Base (KB). In this study, to address the challenges of global entity disambiguation, we proposed Conditional Masked Entity Model Using Knowledge Graph Regularization (CMEM-KG), based on a conditional masked language model, in which multiple mentions in a context can be disambiguated in one forward pass. In addition, to address the long-tailed distribution of global entity disambiguation, we proposed a link prediction regularization, in which the entity embeddings were jointly learned to predict knowledge graph links to prevent the model from overfitting. Compared to other global entity disambiguation models, the model proposed in this study exhibited improved performance on six public datasets without an iterative decoding.