A Dictionary to Translate Change Tasks to Source CodeKatja Kevic, Thomas FritzDepartment of InformaticsUniversity of Zurich, Switzerland{kevic, fritz}@if i .uzh.chABSTRACTAt the beginning of a change task, software developers spenda substantial amount of their time searching and navigatingto locate relevant parts in the source code. Current ap-proaches to support developers in this initial code search pre-dominantly use information retrieval techniques that lever-age the similarity between task descriptions and the iden-tif i ers of code elements to recommend relevant elements.However, the vocabulary or language used in source codeoften dif f ers from the one used for describing change tasks,especially since the people developing the code are not thesame as the ones reporting bugs or def i ning new features tobe implemented. In our work, we investigate the creation ofa dictionary that maps the dif f erent vocabularies using in-formation from change sets and interaction histories storedwith previously completed tasks. In an empirical analysison four open source projects, our approach substantially im-proved upon the results of traditional information retrievaltechniques for recommending relevant code elements.Categories and Subject DescriptorsD.2.7 [Distribution, Maintenance, and Enhancement]General TermsExperimentationKeywordsDictionary, location, change task, interaction history1. INTRODUCTIONA lot of approaches to support and speed up developers inlocating relevant places in the source code for change tasksare based on information retrieval (IR) techniques (e.g., [15,18]). All of these approaches require a similarity in the termsused in change tasks and in the source code of a project [7].While studies have shown that there is some textual over-lap between the vocabulary used for change tasks and theone used for changed classes [16], these vocabularies candif f er signif i cantly and hamper the ef f ectiveness of these ap-proaches [11, 19]. This variability can stem from the dif-ference in the people developing code and the ones creatingPermission to make digital or hard copies of all or part of this work forpersonal or classroom use is granted without fee provided that copies arenot made or distributed for prof i t or commercial advantage and that copiesbear this notice and the full citation on the f i rst page. To copy otherwise, torepublish, to post on servers or to redistribute to lists, requires prior specif i cpermission and/or a fee.MSR’14 May 31 - June 07 2014, Hyderabad, IndiaCopyright 2014 ACM 978-1-4503-2863-0/14/05 ...$15.00.change tasks, the often large number and f l uctuation of peo-ple that work on a project, as well as the variability in termspeople use to describe the same concept.In our work, we consider the vocabulary used for thechange tasks and the one used for the source code of a projectas two dif f erent languages. Similar to Hindle et al. [13], weassume that source code is a natural language written byreal people. A concept can then be expressed in the sourcecode language (denoted as SCL) as well as in the naturallanguage used in change tasks (denoted as NL), and theproblem of locating relevant code for a change task can beseen as a translation problem of a concept from NL to SCL.In this work, we investigate whether we can automaticallycreate an ef f ective dictionary from NL to SCL and use it totranslate a change task into relevant source code elements.We analyze which mapping between NL and SCL is bestto create the dictionary, either using change sets associatedwith change tasks or interaction histories captured as taskcontexts by Mylyn [4]. While change sets contain the actualcode changes for a task, a task context captures a developer’sinteraction, such as selections and edits, with code elementsin the IDE for a task. Finally, we investigate whether ourdictionary approach is more ef f ective in locating relevantcode elements than a well-established IR technique.In an empirical investigation on four open source projects,we found that a dictionary created from a combination ofchange set and task context information performs the bestover all projects and outperforms the IR approach by 73%on class level and by 450% on method level. The resultsalso indicate that a dictionary approach works better whenthere are more change tasks to create the dictionary fromand that for projects with few change tasks, combining thedictionary with IR techniques would be benef i cial.This paper makes the following research contributions:• A novel approach for concept location by translatingchange tasks to source code elements via a dictionary.• An investigation of dif f erent NL-SCL mappings based onchange sets and task context.• An empirical analysis on 4 OS projects showing its ef f ec-tiveness and improvement over traditional approaches.This work represents an initial step towards an automatictranslation of change tasks to source code with the potentialto enhance existing concept location techniques.2. RELATED WORKResearch related to concept location can be categorizedinto dynamic, static and hybrid approaches. An overviewcan be found in [7]. Dynamic approaches (e.g., [9, 21]) usePermission to make digital or hard copies of all or part of this work for personal orclassroom use is granted without fee provided that copies are not made or distributedfor prof i t or commercial advantage and that copies bear this notice and the full citationon the f i rst page. Copyrights for components of this work owned by others than ACMmust be honored. Abstracting with credit is permitted. To copy otherwise, or republish,to post on servers or to redistribute to lists, requires prior specif i c permission and/or afee. Request permissions from Permissions@acm.org.Copyright is held by the author/owner(s). Publication rights licensed to ACM.MSR’14, May 31 – June 1, 2014, Hyderabad, IndiaACM 978-1-4503-2863-0/14/05http://dx.doi.org/10.1145/2597073.2597095320
