.png){kind=icon}
.png){kind=icon}
Return
Embeddings based on relation-specific constraints for open world knowledge graph completion
.png){kind=icon}
PRE
AI
DOI：10.1007/s10489-022-04247-z.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
In this paper, we look at the problem of completing an open-world knowledge graph by describing entities. Many open-world knowledge graph completion (KGC) models train alignment functions that map embeddings based on textual descriptions to structural embedding spaces. They then use traditional closed-world KGC models to predict missing facts. When generating and aligning textual embeddings for the open world knowledge graph completion task, most existing approaches ignore the effect of relations. This means that noise from the texts gets into feature extraction, which hurts model performance. So, a new open-world KGC model called EmReCo is proposed. It is based on relation-specific constraints and focuses on relation-correlated information extraction from entity descriptions with a relation-aware attention aggregator for better textual embeddings. A relation-specific gate filtering mechanism is also made to keep relation-specific features in the entity embeddings. Extensive tests on two benchmark open-world datasets show that EmReCo gets great results, especially with long text datasets, and that the best metric, Hits@1, can get better by 8.1%.
Keywords:
Knowledge graph completion
Open-world
Link prediction
Attention aggregator
Relation-specific
Gate filtering mechanism