Knowledge Graph Entity Type Prediction with Relational Aggregation Graph Attention Network

Abstract

AbstractMost of the knowledge graph completion methods focus on inferring missing entities or relations between entities in the knowledge graphs. However, many knowledge graphs are missing entity types. The goal of entity type prediction in the knowledge graph is to infer the missing entity types that belong to entities in the knowledge graph, that is, (entity, entity type=?). At present, most knowledge graph entity type prediction models tend to model entities and entity types, which will cause the relations between entities to not be effectively used, and the relations often contain rich semantic information. To utilize the information contained in the relation when performing entity type prediction, we propose a method for entity type prediction based on relational aggregation graph attention network (RACE2T), which consists of an encoder relational aggregation graph attention network (FRGAT) and a decoder (CE2T). The encoder FRGAT uses the scoring function of the knowledge graph completion method to calculate the attention coefficient between entities. This attention coefficient will be used to aggregate the information of relations and entities in the neighborhood of the entity to utilize the information of the relations. The decoder CE2T is designed based on convolutional neural network, which models the entity embeddings output by FRGAT and entity type embeddings, and performs entity type prediction. The experimental results demonstrate that the method proposed in this paper outperforms existing methods. The source code and dataset for RACE2T can be downloaded from: https://github.com/GentlebreezeZ/RACE2T.KeywordsKnowledge graphEntity typeRelational aggregationAttentionScoring functionConvolutional neural network