TransRPRT: A Knowledge Representation Learning Model Based on Relationship Transfer and Relationship Mapping Properties

Yanjun Guo,Junlin Zhou

DOI: https://doi.org/10.1109/icsp62122.2024.10743831

2024-01-01

Abstract:Representing learning models can vectorize entities and relationships through vectorization, and can learn the attributes of entities to complete the knowledge graph, which has become a current research hotspot. Some representation learning models rely on the translation rule of TransE. The performance of these models is not ideal when inferring symmetric, reflexive, and synthetic relationship patterns, as well as complex relationships of 1-N, N-1, and N-N relationships simultaneously. For this purpose, we have proposed a knowledge representation learning model based on relationship transfer and relationship mapping attributes, namely TransRPRT. The core idea of the model is to first use the idea of hyperplane projection to embed entities and relationships; Secondly, adding relationship mapping attributes to enhance the model's ability to handle complex relationships; Then, this article proposes a relationship transfer mechanism in order to weaken translation rules; Finally, use probability method to complete the training task. Experiments on five public datasets, FBl5K237, WN18RR, WN11, FB13, and FB15K, have shown that our model has superior performance in link prediction and triplet classification tasks compared to state-of-the-art baselines, verifying the effectiveness of this model.