Infusing Artificial Intelligence into Requirements Quality Assurance

需求捕获了所提议的系统所需的特征、功能和属性。如果不加以解决，需求中的缺陷可能会波及整个开发过程，可能会导致成本超支、系统构建质量不佳和项目失败。为了尽早减少和识别需求缺陷，需要对需求质量保证(RQA)采取系统的措施。在系统和软件工程中，RQA指的是旨在确保系统需求满足期望的质量属性的过程和活动，例如，完整性和无二义性。对于复杂的系统，完全手动的RQA方法不仅代价高昂，而且容易出错。因此，对RQA的自动化支持是重要的。尽管已有研究，但RQA自动化仍然存在重大挑战。第一组挑战涉及需求文档中普遍使用的自然语言(NL)。NL不容易实现自动化，而且还为质量问题的发生留下了很大的空间。目前，针对自然语言需求的几个关键RQA活动，例如歧义检测和完整性检查，几乎没有自动化的支持。同样，在将自然语言需求转换为可用于模拟和测试目的的模型方面，自动化也很少。第二组挑战来自这样一个事实，即系统越来越需要遵守标准、法律和法规。这使得明确处理法律要求对于最大限度地减少不遵守的风险至关重要。到目前为止，几乎没有针对为确保法律要求的质量提供自动协助的研究。这项研究将为RQA设计新的自动化支持，重点是在工业环境中的适用性。这项研究的主要假设是，人工智能的最新进展可以极大地提高准确性，减少与一些困难的RQA任务相关的工作量。这项研究将采用自然语言处理和机器学习的组合来从自然语言需求中提取结构化信息并对这些信息进行分类。这项研究将进一步利用模型驱动工程来表示和分析从NL需求中提取的结构化信息，以及表征需要检查法律需求质量的法律条款。预计该研究将(1)在质量保证方面节省大量成本，并增强对软件密集型系统可靠性的信心，以及(2)使加拿大行业在开发系统和软件质量改进工具方面具有竞争优势。同样重要的是，这项研究将为培养具有软件工程和应用人工智能深度专业知识的高素质人才提供理想的环境，并能够满足研究、社会和经济的需求。

Requirements capture the desired characteristics, functions, and properties of a proposed system. If left unaddressed, defects in requirements may ripple through the entire development process, potentially leading to cost overruns, poorly built systems, and project failures. To mitigate against and identify requirements defects as early as possible, systematic measures are necessary for requirements quality assurance (RQA). In systems and software engineering, RQA refers to procedures and activities aiming to ensure that the requirements of a system meet the desired quality attributes, for example, completeness and unambiguity. For complex systems, a fully manual approach to RQA would not only be expensive but also error--prone. Automated support for RQA is thus important.  Despite the existing research, major challenges remain in RQA automation. A first set of challenges relates to the ubiquitous use of natural language (NL) in requirements documents. NL does not lend itself easily to automation, and further, leaves ample room for quality issues to occur. Currently, several key RQA activities for NL requirements, for example, ambiguity detection and completeness checking have little automated support. Similarly, automation is scarce for transforming NL requirements into models that can be used for simulation and testing purposes. A second set of challenges is posed by the fact that systems increasingly have to comply with standards, laws, and regulations. This makes an explicit treatment of legal requirements paramount to minimize the risk of non--compliance. To date, little research has been directed at providing automated assistance for ensuring the quality of legal requirements. The research will devise novel automated support for RQA with an emphasis on applicability in industrial settings. The main hypothesis underlying the research is that recent advances in artificial intelligence can dramatically increase the accuracy and reduce the effort associated with some difficult RQA tasks. The research will employ a combination of natural language processing and machine learning for extracting structured information from NL requirements and classifying this information. The research will further capitalize on model-driven engineering for representing and analyzing the structured information extracted from NL requirements as well as for characterizing the legal provisions against which the quality of legal requirements needs be checked.  The research is expected to (1) lead to major cost savings in quality assurance and increased confidence in the dependability of software--intensive systems, and (2) give the Canadian industry a competitive advantage in developing systems and software quality improvement tools. As importantly, the research will provide an ideal context for training highly qualified personnel with in--depth expertise in software engineering and applied AI, and capable of fulfilling the needs of research, society and the economy.