AbstractAt its heart, design is a decision-making process. These decisions, and the reasons for making them, comprise the design rationale (DR) for the designed artifact. If available, DR provides a comprehensive record of the reasoning behind the decisions made during the design. Unfortunately, while this information is potentially quite valuable, it is usually not explicitly captured. Instead, it is often buried in other design and development artifacts. In this paper, we study how to identify rationale from text documents, specifically software bug reports and design discussion transcripts. The method we examined is statistical text mining where a model is built to use document features to classify sentences. Choosing which features are most likely to be good predictors is important. We studied two evolutionary algorithms to optimize feature selection – ant colony optimization and genetic algorithms. We found that for many types of rationale, models built with an optimized feature set outperformed those built using all the document features.

No abstract

Design rationale refers to the explicit documentation of the reasons design decisions are made, the alternative design options considered, and the justification supporting these decisions. If properly recorded during the software development life cycle, design rationale documentation can be extremely useful to software engineers.However, because of the costs associated with maintaining extra documentation, it is often omitted. Fortunately, design rationale is naturally embedded in other documents that are created in the general course of software development -such as bug reports, project specifications, emails, or meeting transcripts. I will introduce the text mining problem of automatic extraction of design rationale from such software documentation, and will describe the machine learning based approach to capturing the rationale. I will discuss the feature selection step of the machine learning method, and give background on the genetic algorithm and ant colony optimization metaheuristic chosen in this research for the feature selection procedure. I will show that by applying these text mining techniques to classification of design rationale in software documentation, we are able to achieve higher F-measure than our estimate of what would be achievable manually.