Research On Link Prediction Based On Triangular Structure Relation Inference

In recent years,the knowledge graph has demonstrated rich application value in various fields such as search and intelligent question answering.However,existing knowledge graphs suffer from the issues of data sparsity and incompleteness,which hinder their application in different domains.Link prediction aims to complete the knowledge graph.Most existing link prediction tasks employ convolutional neural network models,which primarily focus on enhancing the interaction between entities and relations but overlook other factors that need to be considered in link prediction tasks,therefore convolutional neural network models used for link prediction tasks face the following problems:(1)The model only handles direct relations between entity pairs in the knowledge graph,disregarding the rich semantic information contained in multi-step relation paths between entity pairs,which is crucial for knowledge representation and link prediction;(2)In the convolutional neural network model ConvR,the embedding dimensions of entities and relations are different,which can have an impact on downstream tasks.If the embedding dimension of the relation is adjusted to be consistent with the entity,it would lead to significant performance degradation.To address the aforementioned issues,this paper proposes a novel convolutional model based on triangular structure relation inference,called TRI-ConvMR.The proposed model adopts the encoder-decoder structure and focuses on the following research aspects:(1)On the basis of the convolutional neural network model,this paper introduces the encoder model based on triangular structure relation inference,called TRI.The triangular structure consists of three connected entity pairs and the relations between them.TRI utilizes the characteristics of the triangular structure to ensure reliable and computational efficient of relation inference while capturing rich semantic information in relation paths for the convolutional neural network model.(2)This paper maps the embedding vectors of the entity in relation paths to the vector space of relation and incorporates them into the process of relation inference.The improved inference algorithm exhibits strong interpretability and performance enhancement.Additionally,to account for the possibility of multiple relation paths under the same entity sequence,the attention mechanism is employed to comprehensively utilize the information of the path sets.(3)On the basis of the ConvR model,this paper proposes the adaptive convolutional model for improving relational embedding,called ConvMR.ConvMR dynamically adjusts the initial embedding structure of relations using multilayer perceptron or convolutional neural network methods.It solves the issues present in the ConvR model and significantly improves prediction performance.Finally,extensive experiments are conducted on five commonly used link prediction datasets: WN18 RR,FB15k-237,WN18,FB15 k,and kinship datasets.Under unified evaluation metrics and the same experimental environment,this paper completed baseline model comparison experiments,ablation experiments,and parameter rationality experiments.The experiment results validate the effectiveness and rationality of the TRI-ConvMR model and its constituent modules.