REquirements TRacing On target (RETRO) enhanced with an automated thesaurus builder: An empirical study

Abstract

Several techniques have been proposed to increase the performance of the tracing process, including use of a thesaurus. Some thesauri pre-exist and have been shown to improve the recall for some datasets. But the drawback is that they are manually generated by analysts based on study and analysis of the textual artifacts being traced. To alleviate that effort, we developed an application that accepts textual artifacts as input and generates a thesaurus dynamically, we call it Thesaurus Builder. We evaluated the performance of REquirements TRacing On target (RETRO) with a Thesaurus generated by Thesaurus Builder. We found that recall increased from 81.9% with no thesaurus to 87.18% when the dynamic thesaurus was used. We also found that Okapi weighting resulted in better recall and precision than TF-IDF weighting, but only precision was statistically significant.