Improvement of Web Semantic and Transformer-Based Knowledge Graph Completion in Low-Dimensional Spaces

Abstract

In recent years, knowledge graph completion (KGC) has garnered significant attention. However, noise in the graph poses numerous challenges to the completion of tasks, including error propagation, missing information, and misleading relations. Many existing KGC methods utilize the multi-head self-attention mechanism (MHA) in transformers, which yields favorable results in low-dimensional space. Nevertheless, employing MHA introduces the risk of overfitting due to a large number of additional parameters, and the choice of model loss function is not comprehensive enough to capture the semantic discriminatory nature between entities and relationships and the treatment of RDF indicates that the dataset contains only positive (training) examples, and the error facts are not encoded, which tends to cause overgeneralization.