A Multidocument Summarization Technique for Informative Bug Summaries

Abstract

To help developers grasp bug information, bug summaries should contain bug descriptions and information on the reproduction steps, environment, and solutions to be informative for developers. However, previously established bug report summarization techniques generate bug summaries mainly by identifying significant sentences, which often miss those bug report attributes. In this paper, we aim to generate informative summaries that cover these specific bug report attributes in a structured form. There are two challenges. First, the relevant information is sometimes scattered over multiple sources. Second, information on the reproduction steps and environment is often filtered out by previous techniques, which identify significant sentences on the basis of their relationships. Therefore, we propose a bug summarization technique that collects information from multiple sources, including duplicates and pull requests, and a classification technique for identifying sentences that provide relevant information on the reproduction steps and environment. Our proposed technique, ClaSum, consists of four steps: preprocessing, classification, sentence ranking, and summarization. We adopted RoBERTa for our classification step, Opinion and Topic association scores for the sentence ranking step, and BART for the summarization step. Our comparative experiments show that our technique outperforms the state-of-the-art technique BugSum in terms of the F1 score by 14%, 8%, and 35% on the SDS, ADS, and DDS datasets, respectively. Additionally, our qualitative investigation shows that our technique generates a more structural summary than two well-known LLMs, Gemini and Claude.