Abstract: To break the limitations brought by existing process mining algorithms that could not be used when logs were missing,a novel method for deep automatic generation of business process models from process text descriptions was proposed based on the existing deep learning and natural language processing technology base.The existing named entity method was improved,and the activity entity recognition model was constructed by Bidirectional Encoder Representation from Transformers (BERT),Bi-directional Long Short Term Memory (BiLSTM),Conditional Random Fields (CRF),and the business process-oriented activity entity recognition method was proposed.The language model was extended from sentence level to document level,and a recursive architecture Ordered Neurons LSTM (ON-LSTM) was proposed to unsupervised discover the activity embedded in the process description document The hierarchical tree was finally transformed into a business process model by using the principle of hierarchical depth of active entities.Experiments were conducted on 150 real System Applications and Products (SAP) product user guide texts collected and labeled manually as training data,and several group experiments were conducted on the basis of ON-LSTM using K-fold cross-validation idea,which verified the effectiveness of the proposed method.

Key words: deep learning, business process discovery, active entity, hierarchical structure, ordered neurons long short term memory

摘要： 为打破现有过程挖掘算法在日志缺失时无法使用带来的局限性,基于现有的深度学习、自然语言处理技术基础,提出一种新颖的从过程文本描述中深度自动生成业务过程模型的方法。对现有命名实体方法进行改进,通过BERT,BiLSTM,CRF构建活动实体识别模型,提出面向业务过程的活动实体识别方法;将语言模型从句子级别扩展到文档级别,提出一种通过递归体系结构有序神经网络(ON-LSTM)无监督地发现过程描述文档中所蕴含的活动实体间潜在的层次结构;通过活动实体的层次深度原则,将层次结构树转化为业务过程模型。通过对人工采集与标注的150个真实的SAP产品用户指南文本作为训练数据进行实验,并在ON-LSTM基础上采用K折交叉验证思想对数据进行多次分组实验,验证了所提方法的有效性。

关键词: 深度学习, 业务过程发现, 活动实体, 层次结构, 有序神经长短期记忆网络