Controlled Natural Language Development and Evaluation for Automotive Requirements

Abstract

Automotive software development depends on requirements that are readable by engineers and precise enough to support later analysis, implementation, and valida tion. In the Volvo Group context, requirements are maintained in a Requirements Engineering Tool across functional, logical, and software levels. They vary in wording, structure, abstraction level, use of tables, and dependence on surrounding metadata and context. This makes it difficult to restructure requirements consistently into a controlled natural language (CNL) and to determine which requirements can be translated and validated without changing the intended meaning or inventing missing information. This thesis is a design science research study that develops a classification-driven, layered CNL and analysis pipeline for software-level automotive requirements. The artifact consists of a grounded software-level taxonomy, a stable set of recurring semantic patterns with an associated grammar, and a proposed pipeline that sepa rates the main requirement statement from notes and supporting material before classification, translation, and validation. The resulting artifact also makes the scope boundary explicit by distinguishing supported requirement forms from ambiguous, metadata-dependent, or otherwise unsupported cases. This supports earlier and safer review decisions during requirements engineering. The artifact was developed and refined iteratively through requirement review, practitioner input, and successive CNL design cycles. The evaluation combined practitioner feedback with review of unseen requirements. Engineers generally judged the CNL clearer and less ambiguous than the original wording, although some table-heavy rewrites were less preferred. On unseen require ments, the artifact covered 166 of 198 requirements, which is 83.8% raw coverage. When the 24 unsupported requirements are excluded, the in-scope coverage was 95.4% (166 of 174). This shows that most requirements within the intended scope can be translated consistently while unsupported cases can be identified early instead of being forced into unsafe translation. The main limitations are that the study is grounded in one industrial setting and that the full end-to-end automated pipeline has not yet been evaluated in operation.