Requirements elicitation through viewpoint control in a natural language environment

Abstract

While requirements engineering is about building a conceptual model of part of reality, requirements validation involves assessing the model for correctness, completeness, and consistency. Viewpoint resolution is the process of comparing different views of a given situation and reconciling different opinions. In his doctoral dissertation Leite [72] proposes viewpoint resolution as a means for early validation of requirements of large systems. Leite concentrates on the representation of two different views using a special language, and the identification of their syntactic differences. His method relies heavily on redundancy: two viewpoints (systems analysts) should consider the same topic, use the same vocabulary, and use the same rule-based language which constrains how the rules should be expressed. The quality of discrepancies that can be detected using his method depends on the quality of the viewpoints. The hypothesis of this thesis is that, independently of the quality of the viewpoints, the number of viewpoints, the language, and the domain, it is possible to detect better quality discrepancies and to point out problems earlier than Leite's method allows. In the first part of this study, viewpoint-oriented requirements engineering methods are classified into categories based on the kind of multiplicity the methods address: multiple human agents, multiple specification processes, or multiple representation schemes. The classification provides a framework for the comparison and the evaluation of viewpoint-based methods. The study then focuses on the critical evaluation of Leite's method both analytically and experimentally. Counter examples were designed to identify the situations the method cannot handle. The second part of the work concentrates on the development of a method for the very early validation of requirements that improves on Leite's method and pushes the boundaries of the validation process upstream towards fact-finding, and downstream towards conflicts resolution. The Viewpoint Control Method draws its principles from the fields of uncertainty management and natural language engineering. The basic principle of the method is that, in order to make sense of a domain one must learn about the information sources and create models of their behaviour. These models are used to assess pieces of information, in natural language, received from the sources and to resolve conflicts between them. The models are then reassessed in the light of feedback from the results of the process of information evaluation and conflict resolution. Among the implications of this approach is the very early detection of problems, and the treatment of conflict resolution as an explicit and an integral part of the requirements engineering process. The method is designed to operate within a large environment called LOLITA that supports relevant aspects of natural language engineering. In the third part of the study the Viewpoint Control Method is applied and experimentally evaluated, using examples and practical case studies. Comparing the proposed approach to Leite's shows that the Viewpoint Control Method is of wider scope, is able to detect problems earlier, and is able to point out better quality problems. The conclusions of the investigation support the view that underlines the naivety of assuming competence or objectivity of each source of information.