# KG-BERT: BERT for Knowledge Graph Completion

**Authors:** Liang Yao, Chengsheng Mao, Yuan Luo  
**arXiv:** [1909.03193](https://arxiv.org/abs/1909.03193)  
**Date:** 2019-09-07  
**Summary page:** https://www.summarizepaper.com/en/paper/1909.03193v2/

## Key points

- Authors propose KG-BERT framework to address knowledge graph incompleteness
- KG-BERT utilizes pre-trained language models for knowledge graph completion
- Triples in knowledge graphs treated as textual sequences and modeled using BERT
- Scoring function computed by KG-BERT helps determine relevance and accuracy of triples
- Experiments conducted on multiple benchmark knowledge graphs
- KG-BERT achieves state-of-the-art performance in triple classification, link prediction, and relation prediction tasks
- Innovative solution to enhance knowledge graph completeness by leveraging pre-trained language models like BERT

## Summary

In the paper titled "KG-BERT: BERT for Knowledge Graph Completion," authors Liang Yao, Chengsheng Mao, and Yuan Luo propose a novel framework to address the issue of incompleteness in knowledge graphs. Knowledge graphs are essential resources for various artificial intelligence tasks but often lack comprehensive information. To overcome this limitation, the authors suggest utilizing pre-trained language models for knowledge graph completion. The proposed KG-BERT framework treats triples in knowledge graphs as textual sequences and leverages the Bidirectional Encoder Representations from Transformers (BERT) model to model these triples. By taking entity and relation descriptions of a triple as input, KG-BERT computes a scoring function using the KG-BERT language model. This scoring function helps determine the relevance and accuracy of a triple within the knowledge graph. To evaluate the effectiveness of their approach, the authors conduct experiments on multiple benchmark knowledge graphs. The results demonstrate that KG-BERT achieves state-of-the-art performance in tasks such triple classification, link prediction, and relation prediction. Overall, this paper introduces an innovative solution to enhance knowledge graph completeness by leveraging pre-trained language models like BERT. The experimental findings highlight the efficacy of KG-BERT in improving various AI tasks reliant on knowledge graphs.