以目標導向使用案例強化EARS軟體需求

Abstract

在需求工程領域中存在三項挑戰：第一，需求分類在樣本數少的類別上表現不佳，且高度依賴關鍵詞；第二，自然語言需求具有模糊性與歧義性；第三，手動擷取目標導向使用案例模型耗時且需經驗與技術。為了解決這些問題，本研究提出一套自動化的需求分析與轉換流程，可用於從自然語言需求中推導出 EARS 需求與目標導向使用案例模型。

首先，我們設計一種使用集成學習的自動化需求分類流程，整合主題、句子與詞彙三種層級的語意分類，並透過兩階段分群演算法與軟分群以符合需求的語意模糊性，最後利用有序加權平均運算子（OWA）動態調整與整合模型的預測結果。結果顯示在 PROMISE_exp 資料集中，我們的方法在非功能需求中各類別表現皆優於傳統方法，即使關鍵詞被替換，F1-score 仍能維持較佳表現。

其次，我們提出一套從自然語言需求自動轉換為 EARS 需求的流程，包含規則導向的轉換機制，可判別六種 EARS 類別，以及結合語意與語法相似度的評估機制，以確保 EARS 轉換的完整性與一致性。

最後，我們自動化生成目標導向使用案例模型。透過語意角色標註擷取潛在使用案例，進行過濾與合併；接著，使用擴散模型與 PDTB 2.0 資料集，推導需求間的語意關係，對應至「包含」、「擴展」與「目標至使用案例」三種使用案例模型中的關係。將目標分類後，產生目標導向使用案例圖與使用案例規格。我們亦設計一套評估機制，驗證從需求到目標、目標到使用案例、使用案例到測試案例三個層級的覆蓋度與可追蹤性。

In requirements engineering, there are three challenges: (1) requirements classification has worse performance in low proportion subcategories of non-functional requirements (NFRs) and is dependent on the appearance of keywords; (2) natural language (NL) requirements are often vagueness and ambiguous; and (3) manually extracting goal-driven use case models is time-consuming and requires experience and skills. To address these issues, we propose an automated pipeline for analyzing and transforming NL requirements into structured Easy Approach to Requirements Syntax (EARS) and goal-driven use case models.

First, we design an automated requirements classification process using ensemble learning, which integrates semantic classification at the topic, sentence, and word levels. We then apply our two-stage clustering algorithm with soft clustering to capture fuzziness in requirements semantics. Finally, we utilize the Ordered Weighted Averaging (OWA) operator to dynamically adjust and aggregate the predictions from different models. Evaluation results on the PROMISE_exp dataset indicate that our method outperforms the traditional approach in NFR subcategories and maintains a higher F1-score even when keywords are replaced.

Second, we propose an automated process to convert NL requirements into EARS requirements, including a rule-based conversion mechanism for identifying six types of EARS requirements and an evaluation mechanism that examines both semantic and syntactic similarity to help ensure the consistency and completeness of the EARS transformation.

Finally, we automatically generate goal-driven use case models. We extract potential use cases using semantic role labeling, perform filtering and merging, and identify discourse relations between requirements using a diffusion model and the PDTB 2.0 dataset. These relations are mapped to three types of relations in the use case model: include, extend, and from goal to use case. After identifying goals, we obtain the goal-driven use case model and generate the use case diagram and use case specification. To assess the quality of the generated model, we propose a goal-driven evaluation that measures the coverage and traceability from requirements to goals, from goals to use cases, and from use cases to test cases.