Abstract: This paper focuses the problem of learning the knowledge low-dimensional embedding representation for entities and relations extracted from domain-specific datasets. Existing embedding methods aim to represent entities and relations from a knowledge graph as vectors in a continuous low-dimensional space. Different approaches have been proposed, being usually evaluated on standard benchmark knowledge graphs, such as Wordnet and Freebase. However, the nature of such data sources prevents those methods of taking advantage of more detailed and enriched metadata, lacking more accurate results on the evaluation tasks. In this paper, we propose HEXTRATO, a novel embedding approach that extends a traditional baseline model TransE by adding ontology-based constraints in order to better capture the relationships between categorised entities and their symbolic representation in the vector space. Our method is evaluated on an adapted version of Freebase, on a publicly available dataset used on ma chine learning benchmarks, and on two datasets in the clinical domain. Our method outperforms the state-of-the-art accuracy on the link prediction task, evidencing the learnt entity and relation embedding representation can be used to improve more complex embedding models.

 

@inproceedings{Hextrato2018,
author    = {Hegler Tissot},
title     = {{HEXTRATO:} Using Ontology-based Constraints to 
             Improve Accuracy on Learning Domain-specific Entity and 
             Relationship Embedding Representation for Knowledge Resolution},
booktitle = {Proceedings of the 10th International Joint Conference on 
             Knowledge Discovery, Knowledge Engineering and Knowledge Management, 
             {IC3K} 2018, Volume 1: KDIR, Seville, Spain, September 18-20, 2018.},
pages     = {70--79},
year      = {2018},
doi       = {10.5220/0006923700700079}
}