# Enhancing Knowledge Graph Embedding Models with Semantic-driven Loss Functions

**Authors:** Nicolas Hubert, Pierre Monnin, Armelle Brun, Davy Monticolo  
**arXiv:** [2303.00286](https://arxiv.org/abs/2303.00286)  
**Date:** 2023-03-01  
**Summary page:** https://www.summarizepaper.com/en/paper/2303.00286v2/

## Key points

- The paper proposes a novel approach to improve the performance of knowledge graph embedding models (KGEMs) for link prediction tasks.
- Traditional approaches treat the label of a triple as either true or false, but recent works suggest that all negative triples should not be valued equally.
- Semantically valid negative triples might be high-quality negative triples and should be treated differently from semantically invalid ones.
- The authors propose semantic-driven versions for the three main loss functions for link prediction by injecting background knowledge about relation domains and ranges into the loss functions to treat the scores of negative triples differently.
- The authors conduct an extensive and controlled experimental study on three public benchmark KGs underpinned with different schemas to evaluate their proposed approach.
- Their proposed loss functions systematically provide satisfying results, leading to better MRR and Hits@$10$ values while driving KGEMs towards better semantic awareness.
- Incorporating semantic information globally improves KGEMs and should be incorporated into loss functions.
- Since domains and ranges of relations are largely available in schema-defined KGs, their proposed approach is both beneficial and widely usable in practice.

## Summary

The paper titled "Enhancing Knowledge Graph Embedding Models with Semantic-driven Loss Functions" proposes a novel approach to improve the performance of knowledge graph embedding models (KGEMs) for link prediction tasks. KGEMs are widely used in various applications related to knowledge graphs (KGs), and they are trained using loss functions that consider a batch of scored triples and their corresponding labels. Traditional approaches treat the label of a triple as either true or false, but recent works suggest that all negative triples should not be valued equally. In line with this recent assumption, the authors posit that semantically valid negative triples might be high-quality negative triples and should be treated differently from semantically invalid ones. To achieve this, they propose semantic-driven versions for the three main loss functions for link prediction. Specifically, they inject background knowledge about relation domains and ranges into the loss functions to treat the scores of negative triples differently. The authors conduct an extensive and controlled experimental study on three public benchmark KGs underpinned with different schemas to evaluate their proposed approach. The results show that their proposed loss functions systematically provide satisfying results, leading to better MRR and Hits@$10$ values while driving KGEMs towards better semantic awareness. This highlights that incorporating semantic information globally improves KGEMs and should be incorporated into loss functions. Furthermore, since domains and ranges of relations are largely available in schema-defined KGs, their proposed approach is both beneficial and widely usable in practice. Overall, this paper presents a novel approach to enhance KGEMs' performance by considering semantically valid negative triples differently from semantically invalid ones through injecting background knowledge about relation domains and ranges into the loss functions.