基于多跳信息融合的实体对齐模型

Entity Alignment Model Based on Multi-Hop Information Fusion

实体对齐是融合不同来源知识图谱的关键步骤。现有的实体对齐方法主要利用结构信息和名称信息,对实体属性值的利用还不够充分,同时在结构利用方面主要利用一阶邻域的结构进行信息的传递,对距离较远的邻居实体的感知能力不足。针对以上问题,提出一种基于多跳信息融合的实体对齐模型。使用预训练语言模型编码属性值信息,在模型的输入中使用实体的名称信息和属性值信息,将其分别输入到2个通道的编码器中进行信息融合,通过多重注意力机制融合不同距离的实体信息,分别计算出不同信息表示下的距离矩阵,对矩阵融合调整后得出最终对齐结果。在原始和降质后的DBP15K数据集上的实验结果表明,所提模型相比现有的各基线模型总体上得到了更精确的对齐结果,其中Hits@1性能比最优模型分别提高了2.51和5.54个百分点。

Entity alignment is a critical step in knowledge graph fusion from various sources.Existing entity alignment methods primarily take advantage of structural information and entity names whilst ignoring attribute information in most cases.In terms of structure utilization,it primarily utilizes the structure of a first-order neighborhood for information transmission and lacks the perception of distant neighbors.To address these issues,an entity alignment model based on multi-hop information fusion is proposed herein.A pre-trained language model is used to encode the attribute value information.The entity name and attribute values are input into different model encoders to achieve information fusion.The attention mechanism is used to fuse the entity information at different distances.Distance matrices under different information representations are calculated,and the final alignment result is obtained after the matrix is fused and adjusted.Based on experiments with the original and degraded DBP15K datasets,it can be observed that the proposed model achieves more accurate alignment results overall,compared with the baseline models,where the Hits@1 performance is increased by up to 2.51 and 5.54 percentage points over state-of-the-art models.