Answering Follow-up Questions on Bug Reports with Structured Information Retrieval and Deep Learning

Software bug reports reported on bug-tracking systems often lack crucial information for the developers to promptly resolve them, costing companies billions of dollars. There has been significant research on effectively eliciting information from bug reporters in bug tracking systems using different templates that bug reporters need to use. However, the need for asking follow-up questions persists. Recent studies propose techniques to suggest these follow-up questions to help developers obtain the missing details, but there has been little research on answering these follow up questions, which are often unanswered. In this paper, we propose a novel approach that uses CodeT5 in combination with Lucene, an information retrieval technique that leverages the relevance of different bug reports, their components, and follow-up questions to recommend answers. These top-performing answers, along with their bug report, serve as additional context apart from the deficient bug report to the deep learning model for generating an answer. We evaluate our recommended answers with the manually annotated answers using similarity metrics like Normalized Smooth BLEU Score, METEOR, Word Mover's Distance, and Semantic Similarity. We achieve a BLEU Score of up to 34 and Semantic Similarity of up to 64 which shows that the answers generated are understandable and good according to Google's standard and can outperform multiple baselines.