Advaita: Bug Duplicity Detection System (2001.10376v1)

Abstract: Bugs are prevalent in software development. To improve software quality, bugs are filed using a bug tracking system. Properties of a reported bug would consist of a headline, description, project, product, component that is affected by the bug and the severity of the bug. Duplicate bugs rate (% of duplicate bugs) are in the range from single digit (1 to 9%) to double digits (40%) based on the product maturity , size of the code and number of engineers working on the project. Duplicate bugs range are between 9% to 39% in some of the open source projects like Eclipse, Firefox etc. Detection of duplicity deals with identifying whether any two bugs convey the same meaning. This detection of duplicates helps in de-duplication. Detecting duplicate bugs help reduce triaging efforts and saves time for developers in fixing the issues. Traditional natural language processing techniques are less accurate in identifying similarity between sentences. Using the bug data present in a bug tracking system, various approaches were explored including several machine learning algorithms, to obtain a viable approach that can identify duplicate bugs, given a pair of sentences(i.e. the respective bug descriptions). This approach considers multiple sets of features viz. basic text statistical features, semantic features and contextual features. These features are extracted from the headline, description and component and are subsequently used to train a classification algorithm.