從需求建構類別圖:應用Transformer為基礎的機器學習方法

Abstract

過往將軟體工程需求自動構建類別圖的方法，都是使用建立規則和自然語言處 理技術來抓取類別、類別的屬性和方法和類別之間的關係。然而，此類方法會因 為每個人書寫需求方式的不同，而遺漏屬性或是方法等等，並且也沒有辦法完整 的找到所有類別間的關係。 因此，本研究提出了一個自動化的流程，能夠從軟 體 需求完整地辨別出類別、類別的屬性和方法和類別之間的關係，其包含以下四 個 步驟:將自然語言書寫的軟體需求改寫為EARS(Easy Approach to Requirements Syntax)格式;利用Transformer-based的模型以及自然語言處理技術分析需求; 生成 描述類別圖的UML文件;使用開源軟體工具PlantUML產出類別圖。

The previous methods of automatically converting software engineering require- ments into class diagrams involved using rule-based approaches and natural language processing techniques to extract the classes, attributes, methods, and relationships between classes. However, such methods may miss attributes or methods and cannot fully find all relationships between classes due to different writing styles of individ- uals. Therefore, in this research work, we propose an automated process that can identify classes, attributes, methods, and relationships between classes from soft- ware requirements comprehensively with the following four steps: 1. rewrite the software requirements written in natural language into EARS (Easy Approach to Requirements Syntax) format, 2. Use Transformer-based models and natural lan- guage processing techniques to analyze requirements, 3. Generate UML documents to describe class diagrams, 4. Use the open-source tool PlantUML to produce class diagrams.