Context-based retrieval in software development

Abstract

Although software development may include all the activities that result in a software product, from its conception to its realization, here we focus on the process of writing and maintaining the source code. Software development projects have been increasing in size and complexity, requiring developers to cope with a large amount of contextual information during their work. With workspaces frequently comprising hundreds, or even thousands, of artifacts, they spend a considerable amount of time navigating the source code or searching for a speci c source code artifact they need to work with. With the aim of helping developers understand the source code structure and nd what they need, modern Integrated Development Environments (IDEs) provide several features for searching and navigating the source code. But, according to some studies, developers still spend a considerable amount of time searching and navigating the source code structure. With regard to search, the most commonly used approach is the matching of speci c patterns in the lines of code that comprise a software system, requiring a direct correspondence between the pattern and the text in the source code. The limitations of this approach have been surpassed by the research carried out in the eld of Information Retrieval (IR), encouraging researchers to use these techniques to help developers nding relevant source code for their current task. But, despite the fact that context is argued to improve the e ectiveness of IR systems, as far as we know, none of the previous approaches have used the contextual information of the developer to improve the retrieval or ranking of relevant source code in the IDE. Another interesting form of delivering relevant source code artifacts to a developer is using a recommender system. These systems have been used in a wide variety of domains, to help users nd relevant information, deal with information overload and provide personalized recommendations of very di erent kinds of items. In software development, researchers studied ways of using contextual information to recommend source code artifacts that are potentially relevant for the current task of the developer. But, these approaches usually use a limited context model, or require contextual information to be explicitly provided. The research described in this thesis is focused on the development of a contextbased approach to search and recommendation of source code in the IDE. The source code structure stored in the workspace of the developer is represented in a knowledge base. A context model represents the source code elements that are more relevant for the developer in a speci c moment. These structures are then used to improve the retrieval and ranking of source code elements, such as classes, interfaces and methods, taking into account their relevance to the current context of the developer. The relevance of the source code elements retrieved is computed based on the structural and lexical relations that exist between these elements and the elements in the context model. We have implemented a prototype that implements and integrates our approach in the Eclipse IDE. This prototype was tested with a group of developers in order to validate our approach. The statistical information collected shows that the source code elements manipulated by the developer are highly related. This supports our claim that the relations that exist between source code artifacts can be used to measure the proximity between these artifacts, and to compute their relevance in the current context of the developer. Also, we have veri ed that the context components have a clear contribution to improve the ranking of search results, with the search results selected by the developers using our approach being better ranked in more than half of the times. With respect to recommendations, although the results are not so evident, we have shown that our context model can be used to retrieve relevant source code elements for the developer, being able to predict the needed source code element in more than half of the times.