TAG: UML Activity Diagram Deeply Supervised Generation from Business Textural Specification

Unified modeling language (UML) activity diagrams depict the internal behavior of different program operations with the help of nodes and edges, describing the business logic in user requirements. Traditionally, requirements engineers and practitioners refer to business process documents and analyze them to build UML activity diagrams manually, which makes labor and time consuming. Recently, deep learning technology has been utilized in various fields and has achieved excellent results. We propose a novel pipeline, named TAG, for automatically generating UML activity diagrams based on deep learning. The inspiration for TAG is as follows: (1) Semantic roles¹, such as signal and condition entities in texts, can be obtained via sequence labeling; (2) A business process document corresponds to a semantic role sequence. According to the predefined rules, the semantic role sequence is used to construct the graph neural network, and the temporal activity relationship in a business process document is predicted via multi-layer semantic fusion; (3) Use temporal activity relationships to generate UML activity diagrams automatically. The entire process was automatically completed. SAP is the largest non-American software company by revenue. We obtained the original data from the SAP website and sorted them as business process documents. After preliminary experiments, a temporal activity relationship prediction accuracy rate of 79.87% was achieved. Simultaneously, some business process documents are available from https://github.com/lwx142857/bussiness-process.