Towards the automatic classification of traceability links

Abstract

A wide range of text-based artifacts contribute to software projects (e.g., source code, test cases, use cases, project requirements, interaction diagrams, etc.). Traceability Link Recovery (TLR) is the software task in which relevant documents in these various sets are linked to one another, uncovering information about the project that is not available when considering only the documents themselves. This information is helpful for enabling other tasks such as improving test coverage, impact analysis, and ensuring that system or regulatory requirements are met. However, while traceability links are useful, performing TLR manually is time consuming and fraught with error. Previous work has applied Information Retrieval (IR) and other techniques to reduce the human effort involved; however, that effort remains significant. In this research we seek to take the next step in reducing it by using machine learning (ML) classification models to predict whether a candidate link is valid or invalid without human oversight. Preliminary results show that this approach has promise for accurately recommending valid links; however, there are several challenges that still must be addressed in order to achieve a technique with high enough performance to consider it a viable, completely automated solution.