Hierarchical Embedding Methods for Large-scale Complex Knowledge Graphs

Abstract

Abstract In large-scale complex knowledge graphs (KGs) for knowledge reasoning, most existing methods ignore the hierarchical features of KGs, limiting their efficiency in recommender systems, information retrieval, and intelligent Q&A systems by their discrete nature. Based on the translation-based knowledge graph embedding method, this paper proposes a hierarchical embedding method for large-scale complex KGs, which encodes the semantics of entity representations through continuous bag-of-word (CBOW) and convolutional neural network (CNN) models. At the same time, the hierarchies are used as regularization terms to constrain the learning of entity embeddings. After that, the model joint loss function is designed to combine the descriptive loss function of the entity and the hierarchical loss function through the equilibrium coefficients, and the model optimal equilibrium coefficients are determined adaptively. Finally, we conducted experiments on the benchmark dataset and the constructed domain dataset, and the experimental results show that our method can capture meaningful hierarchical information, which is superior to representative knowledge graph reasoning methods and improves the performance of link prediction.