Mean dependency length — a new metric for requirements quality

Abstract

This paper proposes the mean dependency length (MDL) as a metric for measuring natural language requirements quality. Dependency length is a linguistic feature based on dependency grammar, which natural language researchers have traditionally used to evaluate syntactic complexity in other contexts. In this study, aided by MATLAB-based algorithms, the authors assessed MDL over a requirements set composed of 249 original statements, rephrased into five pattern systems. Null hypothesis and effect size testings revealed that MDL is sensitive to the application of pattern rules and to the differences among the patterns, both in an absolute approach and in comparison with other metrics. Furthermore, it was also demonstrated that MDL is aligned with users' values, especially for understandability issues, and can be measured automatically. Finally, the work concluded that MDL is a convenient metric for assessing the quality of natural language requirements.