Abstract: The rapid development of large language models has a significant impact on business process management in the enterprise domain,leading to improved efficiency,cost reduction,enhanced customer experience,and fostering innovation.The automatic generation of business process in Business Process Management(BPM) is of great significance,allowing for business improvement through simulating and visualizing complex business process.To help improve processes and enhance efficiency,a method for automatic business process generation was proposed to integrate into practical business scenarios.The business process text was transformed into an abstract label sequence using a signal list.Then,a prompt template was constructed to obtain the adjacency table of abstract labels from the large language model,thereby determining the connection relationships between abstract labels and obtaining an initial graph,and the initial graph was inputted into a graph neural network with a transductive learning approach for supervised training.Finally,the direct temporal relationships between activities were predicted and transformed into a process diagram.Experiments showed that the proposed method achieved an overall F1-score of 0.67 in predicting temporal relationships between activities,which outperformed baseline methods and large language model methods in predicting successor,parallel and unrelated temporal relationships,and surpassed baseline methods in predicting selection relationship but fell behind large language model methods.

Key words: large language model, business process management, automatic generation of business process, graph neural network

摘要： 大语言模型的迅速发展对企业领域的业务过程管理产生了提高效率、降低成本、增强客户体验和促进创新等显著影响。业务过程管理(BPM)中的业务过程自动生成具有模拟业务过程进行业务改进以及将复杂的业务过程可视化等重大意义。所提出的业务过程自动生成方法能够整合到实际业务场景中,以帮助改善业务过程并提高效率。所提方法分为以下几个部分,首先将业务过程文本经过信号词库转化为抽象标签序列,其次构建提示模板从大语言模型中得到抽象标签的邻接表从而确定抽象标签之间的连接关系得到一张初始图,随后将初始图输入到归纳式图神经网络进行监督学习训练,最后预测出活动间直接时序关系并将其转化为过程图。实验表明,所提方法在预测活动间时序关系的总体F1-分数达到了0.67,在预测顺序、并发和无关系的时序关系上领先基线方法和大语言模型的方法,在选择关系上能够领先基线方法但落后于大语言模型的方法。

关键词: 大语言模型, 业务过程管理, 业务过程自动生成, 图神经网络