Research On Knowledge Representation Learning Methods For Sparse Entities

In the era of big data,knowledge graph(KG)is an important approach of representing and organizing massive knowledge.Distributed knowledge representation learning(KRL)is a new field in KG.It represents the semantic information of the entities and relations with a dense low dimensional real-valued vector.However,the problem of entity sparsity is common in most knowledge graphs.It leads to the fact that the existing distributed KRL methods cannot learn the representation of sparse entities well.Therefore,this thesis carries out the following work.(1)The activation mechanism is proposed to capture sparse entity features,aiming to enhance existing KRL models.The activation mechanism captures the interaction features between a relation and an entity through the activation vector,and then activates the relationspecific feature dimension of the sparse entity under a certatin relation,which can help the model to capture more the interaction feature dimension of the sparse entity.Experiments show that the activation mechanism can effectively capture the features of sparse entities and improve the link prediction accuracy for sparse entities.In addition,the adaptive negative sample generation method is proposed to reduce the probability of generating false negative triplets during training.It improves the accuracy of triplet classification.(2)The joint KRL method based on relation hierarchical structure and triplet structure,is proposed to alleviate the impact of entity sparsity by introducing relation hierarchical structure information.Most of the existing KRL methods learn entity representation from triplet structure,ignoring additional information.Only triplet structure may not provide accurate semantic representation in sparse knowledge graphs.In this thesis,a joint KRL framework is purposed to learn the representation of relation hierarchical structure and triplet structure respectively.It can jointly enhance the semantic representation of sparse entities.Based on the above work,a joint KRL model(ACTrans H-RHS)that combines activation mechanism and relation hierarchical structure is proposed,to perform KG completion tasks on the scenario of COVID-19 knowledge graph.The experimental results show that ACTrans H-RHS is highly competitive compared with current advanced models in the relation prediction task on COVID-19 knowledge graph.