Preview

PDF Pending copyright assessment - Request a copy 1Mb

Knowledge graph embedding (KGE) models optimize loss functions to maximize the total plausibility of positive triples and minimize the plausibility of negative triples. Negative samples are essential in KGE training since they are not as observable as positive samples. Currently, most negative sampling methods apply different techniques to keep track of negative samples with high scores that are regarded as quality negative samples. While, we found entities with similar semantic contexts are easier to be deceptive and misclassified, contributing to quality negative samples. This is not considered in most negative sampling approaches. Besides, the unequal effectiveness of quality negative samples in different loss functions is usually ignored. In this paper, we propose an Entity Similarity-based Negative Sampling framework (ESNS). The framework takes semantic similarities among entities into consideration with a shift-based logistic loss function. Comprehensive experiments on the five benchmark datasets have been conducted, and the experimental results demonstrate that ESNS outperforms the state-of-the-art negative sampling methods in the link prediction task.

Item Type:	Refereed Conference Paper
Keywords:	knowledge graph, knowledge representation, semantic web
Research Division:	Information and Computing Sciences
Research Group:	Artificial intelligence
Research Field:	Knowledge representation and reasoning
Objective Division:	Information and Communication Services
Objective Group:	Information systems, technologies and services
Objective Field:	Artificial intelligence
UTAS Author:	Yao, N (Ms Naimeng Yao)
UTAS Author:	Li, X (Ms Xiang Li)
UTAS Author:	Bai, Q (Dr Quan Bai)
ID Code:	154217
Year Published:	2022
Deposited By:	Information and Communication Technology
Deposited On:	2022-11-12
Last Modified:	2023-01-11
Downloads:	0