Abstract: Semantic similarity matching of model elements is a key technology for constructing semantic mapping rules and realizing Heterogeneous Model(HM)transformation.The accuracy and objectivity of semantic matching for HM model elements has always been a difficult problem.Current methods exhibit a strong dependence on data amount and expert experience,with high uncertainty and subjectivity in matching models,lacking objective methods for effective matching.To address this,the semantic similarity matching method for model elements based on Graph Attention Networks(GAT)was proposed.The heterogeneous graph networks were constructed for HM graph data format representation.Node and edge feature representations were embedded in each model's graph to obtain corresponding node and edge embedding vectors.The semantic similarity of HM model elements was computed by GAT.Finally,taking the Systems Modeling Language(SysML)state machine metamodel and timed-automata metamodel as example,the effectiveness and rationality of the proposed method was validated.Compared to the existing methods,the proposed method could significantly reduce the dependency on pre-trained data and expert experience.It also lowered the uncertainty and subjectivity of matching models,as well as the cost and difficulty of semantic matching.

Key words: heterogeneous model, model driven architecture, graph attention networks, semantic similarity, model transformation

摘要： 模型元素语义相似度匹配是构建语义映射规则、实现异构模型(HM)转换的关键技术。HM模型元素语义匹配的准确性及客观性一直是个难点问题,目前现有方法对数据量及专家经验的依赖度较强,匹配模型不确定性与个性化程度高,缺乏客观方法进行有效匹配,为解决该问题,提出基于图注意力网络(GAT)的模型元素语义相似度匹配方法。首先,构建异构图网络,实现HM的图数据格式表示;其次,对各模型所表示的图进行节点及边的特征表示嵌入,得到节点嵌入向量以及边嵌入向量;之后,采用GAT进行HM模型元素语义相似度计算;最后,以系统建模语言(SysML)状态机图元模型及时间自动机元模型为例进行验证,证明了所提方法的有效性及合理性。与现有方法进行对比分析,所提方法可以较大程度地减少对预先训练数据及专家经验的依赖度,同时降低匹配模型的不确定性及个性化程度,以及降低语义匹配的成本及实现难度。

关键词: 异构模型, 模型驱动体系架构, 图注意力网络, 语义相似度, 模型转换