Knowledge Representation Learning Based On Fine-grained Entity Description Information And AcrE Model

The knowledge graph contains a large number of entities and relationships,and the relationships among them are complex and require appropriate technical means to represent them so that they can be flexibly applied to other knowledge acquisition tasks.The representational capability of entity and relationship vector representation has an important fundamental impact on the performance of the relevant knowledge acquisition task models.In this paper,we aim to build a knowledge representation learning model with good performance from the perspective of enriching the semantic information contained in entity and relational vector representations,and conduct a theoretical and experimental study on the problem of low differentiation of entity description vector representations in specific semantic scenarios and the problem of excessive model parameters due to the gate mechanism.First,a fine-grained entity description information extraction method based on the attention mechanism is proposed to address the problem of low differentiation of entity description vector representations in specific semantic scenarios.The method uses the BERT pre-training model as the encoding layer to distinguish the semantics of the head entity in specific semantic scenarios by adding relationship and tail entity semantic features to the head entity vector representation,and feature extraction of entity description information by the attention mechanism to alleviate the problem of low differentiation of entity description vector representation in specific semantic scenarios.Secondly,an information fusion method based on parameter sharing is proposed to address the problem of excessive amount of model parameters caused by the gate mechanism.The method extracts features from the entity structure information and the fine-grained entity description information by sharing the parameter matrix between them,so as to alleviate the problem of excessive model parameters caused by the gate mechanism;then,the weight of entity structure information is calculated by normalizing the proportion of entity structure information in the total information(the sum of entity structure information and fine-grained entity description information);finally,the entity structure information and the fine-grained entity description information are weighted and summed to obtain a fused vector representation of the entity,which can effectively enhance the expression capability of the entity vector representation.Finally,based on the above study,a knowledge representation learning model FGDAcr E based on fine-grained entity description information and Acr E model is constructed,and link prediction experiments are conducted on four datasets,WN18,WN18 RR,FB15K and FB15K-237.The experimental results show that the model outperforms other comparison models in two metrics,MR and Hits@10,and validate the effectiveness of the fine-grained entity description information extraction method based on attention mechanism and the information fusion method based on parameter sharing for model performance improvement.