基于子图特征融合的链接预测方法

Link Prediction Method Based on Sub-graph Feature Fusion

链接预测旨在预测知识图谱查询过程中缺失的事实三元组,通常广泛应用于智能问答、信息检索等任务,但由于知识图谱中节点和关系数量庞大,将整个图谱进行编码需要耗费相当大的资源,且图嵌入的编码方式缺少询问句中自带的语义信息,使链接预测结果并不理想。为此,提出一种基于子图嵌入的实体链接方法 LPBS,基于强化学习模型设计相关策略来获取预测链接路径上下文集合并进行输入编码,然后通过基于多头自注意力机制的双塔模型获取询问句和子图的嵌入特征,最后通过交叉注意力机制将量特征融合后得到各节点的预测分布。在自建工业领域数据集上的测试发现,所提方法评的MMR达到0.362,Hits@1达到0.313,并通过消融实验证明了模型的有效性。

Link prediction aims to predict missing fact triplets in the knowledge graph query process,and is commonly used in tasks such as intelligent question answering and information retrieval.However,due to the large number of nodes and relationships in the knowledge graph,encoding the entire graph requires significant resources,and the encoding method of graph embedding lacks the semantic information inherent in the query sentence,resulting in unsatisfactory link prediction results.To this end,a subgraph embedding based entity linking method LPBS is proposed.Based on reinforcement learning models,relevant strategies are designed to obtain the upper and lower text sets of predicted link paths and merge them for input encoding.Then,the embedding features of query sentences and subgraphs are obtained through a dual tower model based on multi head self attention mechanism.Finally,the quantitative features are fused through cross attention mechanism to obtain the predicted distribution of each node.Testing on a self built industrial dataset found that the proposed method achieved an MMR of 0.362,Hits@1 reached 0.313 and demonstrated the effectiveness of the model through ablation experiments.