Text-To-Text Generation for Issue Report Classification

We present our participation in the issue report classification tool competition at the 3rd International Workshop on Natural Language-based Software Engineering. Given the substantial influx of issue reports in large scale software development each day, there is a pronounced need for a system capable of automatically classifying issue reports into categories such as bugs, enhancements, and features. We propose a text-to-text generation-based supervised approach for issue report classification. We fine-tuned and evaluated our approach on the provided dataset of 3000 labeled issue reports (as bugs, enhancements, and questions). Our approach yielded an average cross-repo F1-score of 0.8297 across all classes, which is comparable to the SetFit baseline of 0.8270.CCS CONCEPTS• Software and its engineering → Software verification and validation; • Computing methodologies → Natural language generation; Supervised learning.