化妆品风险物质知识图谱构建及应用

Construction and application of knowledge graph forcosmetic risk substances

针对国内外化妆品风险物质多语言特性和复杂关联的特点,提出一种基于双通道图神经网络的邻域匹配算法。采用图神经网络学习实体属性特征和跨域交互特性,将不同特性实体映射到相同的向量空间,通过邻域匹配网络聚合实体邻域特征,为每个实体构建邻域网络以实现实体对齐,并应用于多语言风险物质知识图谱及问答系统构建。实验结果表明,在化妆品风险物质数据集上该方法获得的Hits@1、Hits@10与MRR值都优于其它基线模型,分别平均提升6.37%、8.17%与9.37%。

Aiming at the multi-lingual characteristics and complex association of domestic and foreign cosmetic risk substances,a neighborhood matching algorithm based on two-channel graph neural network coding was proposed.To map entities with diffe-rent properties to a unified vector space,the graph neural network was proposed to learn entity attributes and cross-domain inte-raction characteristics.The neighborhood matching network was employed to aggregate the neighborhood characteristics of entities,and the neighborhood network was constructed for each entity to achieve entity alignment.It was applied to construct multi-lingual risky substance knowledge graph and question and answer system.Experimental results show that the Hits@1,Hits@10 and MRR values obtained using this method on the cosmetic risk substances dataset are better than those obtained using other baseline models,with an average increase of 6.37%,8.17%and 9.37%,respectively.