Abstract: Traditional knowledge graph embedding models primarily focused on the structural information within triples, without fully leveraging external semantics to enhance the embedding representation capabilities. For instance, they did not adequately consider the multi-step relational path information between entities or the importance of different paths, nor did they utilize entity description information to enhance context-awareness. To improve the effectiveness of knowledge graph applications, a knowledge graph embedding model named MPDRL, was proposed, which integrated multi-step relational paths and entity description information. Firstly, the model encoded the path information between two entities and calculated the path weights, using a self-attention mechanism to obtain the representation of relational path information. Secondly, it employed the BERT (Bidirectional Encoder Representations from Transformers) model to encode entity description information, using bidirectional attention mechanisms to calculate the attention weights between entity description embeddings and triple relation embeddings, thereby enhancing the semantic information of entities. Finally, the model fused the relational path embeddings, entity description embeddings, and triple structure embeddings for joint training. To evaluate the performance, link prediction and triple classification experiments were conducted on public datasets for the proposed model and benchmark models. In the link prediction task on the FB15k-237 dataset, compared with the entity-relation fusion embedding model PDRL (Relation Path and Entity Description Information Based Knowledge Graph Representation Learning, PDRL), the multi-hop path model Att-ConvBiLSTM, and the TPKGE model, the model achieved a 5.7, 2.9, and 2.5 percentage point increase in the Hit@10 metric, respectively. In the triple classification task, the proposed model also improves the accuracy on the dataset by 2.81 and 0.9 percentage points, compared to the best-performing PDRL model. The experimental results demonstrate the feasibility and effectiveness of this model.

Key words: knowledge representation, relational path, knowledge graph, link prediction, text description

摘要： 传统的知识图谱表示学习模型主要聚焦于三元组内部的结构信息，未充分利用外部语义增强嵌入表征能力，如没有充分考虑实体间的多步关系路径信息以及不同路径的重要程度，没有利用实体描述信息增强上下文感知能力。鉴于此，为提升知识图谱应用效果，提出了融合多步关系路径和实体描述信息的知识图谱表示学习模型MPDRL。首先，通过对两实体间的路径信息进行编码，并使用自注意力机制计算路径权重，以获得关系路径信息表示；其次，使用BERT(Bidirectional Encoder Representations from Transformers)模型对实体描述信息进行编码，利用双向注意力机制计算实体描述信息嵌入与三元组关系嵌入之间的注意力权重，增强实体的语义信息；最后，将关系路径信息嵌入、实体描述信息嵌入和三元组结构嵌入进行融合训练。为评估性能，在公开数据集上针对所提模型和基准模型进行了链接预测和三元组分类的实验任务。链接预测任务中，与实体关系融合嵌入模型PDRL(Relation Path and Entity Description Information Based Knowledge Graph Representation Learning, PDRL)、多跳路径模型Att-ConvBiLSTM以及TPKGE模型相比，所提模型在FB15k-237数据集上的Hit@10指标分别提高了5.7、2.9、2.5个百分点；在三元组分类任务上，所提模型在数据集上的准确率较于最优的PDRL模型分别提升了2.81、0.9个百分点。实验结果表明本模型的可行性和有效性。

关键词: 知识表示, 关系路径, 知识图谱, 链接预测, 文本描述