Knowledge Graph Representation Learning via Generated Descriptions

Abstract

Knowledge graph representation learning (KGRL) aims to project the entities and relations into a continuous low-dimensional knowledge graph space to be used for knowledge graph completion and detecting new triples. Using textual descriptions for entity representation learning has been a key topic. However, the current work has two major constraints: (1) some entities do not have any associated descriptions; (2) the associated descriptions are usually phrases, and they do not contain enough information. This paper presents a novel KGRL method for learning effective embeddings by generating meaningful descriptive sentences from entities’ connections. The experiments using four public datasets and a new proposed dataset show that the New Description-Embodied Knowledge Graph Embedding (NDKGE for short) approach introduced in this paper outperforms most of the existing work in the task of link prediction. The code and datasets of this paper can be obtained from GitHub ( https://github.com/MiaoHu-Pro/NDKGE .)