Neighborhood selective aggregation zero-shot knowledge graph link prediction method with single sample support
In order to solve the problem of performance degradation of zero-shot knowledge graph link prediction model under the condition of limited support samples,this paper proposed a neighborhood selective aggregation zero-shot knowledge graph link prediction method with single sample support(NSALP).The method contained three modules,such as feature extractor,generator and discriminator.It improved the feature extractor module by referring to the idea of graph isomorphic network,and assigned a learnable parameter to each neighborhood node when aggregating head and tail neighborhoods,so as to filter irrele-vant features and highlight effective features.The combination of head node embedding and relation text description was used as the guide of the learning process of the generator,so that the new combination features generated by the generator were clo-ser to the real knowledge triple structure features.On NELL-ZS and Wiki-ZS zero-shot knowledge graph datasets,the perfor-mance of the proposed model is improved by 2.5 and 0.7 percentage points respectively compared with the baseline model.In the ablation experiments conducted on NELL-ZS,the performance of the proposed extractor+and generator+modules is better than that of the model without improvement,which proves the effectiveness of the proposed improved method.
万方数据
维普
