On the relationship between similar requirements and similar software

Abstract Recommender systems for requirements are typically built on the assumption that similar can be used as proxies to retrieve software. When a stakeholder proposes new requirement, natural language processing (NLP)-based similarity metrics exploited existing requirements, and in turn, identify previously developed code. Several NLP approaches computation between available. However, there is little empirical evidence their effectiveness code retrieval. This study compares different approaches, from lexical ones semantic, deep-learning techniques, correlates among with of associated The evaluation conducted real-world two industrial projects railway company. Specifically, most pairs across automatically identified using six models. Then, trace links software each pair. then computed JPLag. Finally, correlation evaluated see which model shows highest thus more appropriate In addition, we perform focus group members company collect qualitative data. Results show moderately positive similarity, pre-trained deep learning-based BERT preprocessing outperforming other Practitioners confirm generally regarded proxy similarity. they also highlight additional aspect comes into play when deciding reuse, e.g., domain/project knowledge, information coming test cases, links. Our work first explore relationship quantitative standpoint. useful not only recommender but engineering tasks relevant, such tracing change impact analysis.