| Summary: | In large-scale software development environments, defect reports are maintained through bug tracking systems (BTS) and analyzed by domain experts. Different users may create bug reports in a non-standard manner and may report a particular problem using a particular set of words due to stylistic choices and writing patterns. Therefore, the same defect can be reported with very different descriptions, generating non-trivial duplicates. To avoid redundant work for the development team, an expert needs to look at all new reports while trying to label possible duplicates. However, this approach is neither trivial nor scalable and directly impacts bug fix correction time. Recent efforts to find duplicate bug reports tend to focus on deep neural approaches that consider hybrid representations of bug reports, using both structured and unstructured information. Unfortunately, these approaches ignore that a single bug can have multiple previously identified duplicates and, therefore, multiple textual descriptions, titles, and categorical information. In this work, we propose SiameseQAT, a duplicate bug report detection method that considers information on individual bugs as well as information extracted from bug clusters. The SiameseQAT combines context and semantic learning on structured and unstructured features and corpus topic extraction-based features, with a novel loss function called <italic>Quintet Loss</italic>, which considers the centroid of duplicate clusters and their contextual information. We validated our approach on the well-known open-source software repositories Eclipse, NetBeans, and Open Office, comprised of more than 500 thousand bug reports. We evaluated both the retrieval and classification of duplicates, reporting a Recall@25 mean of 85% for retrieval and 84% AUROC for classification tasks, results that were significantly superior to previous works.
|
|---|
