Knowledge Graph Representation Learning Model Based on Capsule Network and Information Fusion

• Published in: Journal of Applied Science and Engineering Vol. 29; no. 1; pp. 89 - 101
• Main Authors: Chu Zhao, Gilja So, Rui Chen
• Format: Journal Article
• Language: English
• Published: Tamkang University Press 2026
• Subjects: capsule network; information fusion; knowledge representation learning; multi-layer attention mechanism
• ISSN: 2708-9967; 2708-9975
• DOI: 10.6180/jase.202601_29(1).0009

In recent years, knowledge representation learning has played a key role in intelligent recommendation, intelligent question-answering, and intelligent retrieval, and has been widely concerned. Knowledge representation learning aims to vectorize semantic information and deduce knowledge through mathematical formulas with the help of low dimensional embedding of entity and relation. Although knowledge representation learning based on knowledge graph can obtain entity structure and relational embedding, it lacks semantic information utilization of entity description text. In addition, with the increase of the scale of the knowledge graph, the categories and quantities of entities and relationships, as well as the content and sources of entity descriptions, the correspondence between the textual descriptions of entities and the triplet structure information becomes more difficult to obtain. Therefore, we propose a novel knowledge graph representation learning model based on capsule network and information fusion in this paper. Based on anchor node and neighbor node and the relational sampling strategy, each node on the knowledge graph is represented by the predicted operator graph. The capsule network is used to gather the image features for each node to obtain the node representation vector, which is finally input to the decoder to calculate the score. In particular, we construct a loss function for the multi-layer attention mechanism of entity structure and semantic fusion. Experimental results show that the proposed method can effectively deduce the hidden link relationship between entities containing complex entity descriptions, and has more accurate classification accuracy than other methods in triplet classification tasks.