Exploring Developers Discussion Forums for Quantum Software Engineering: A Fine-Grained Classification Approach Using Large Language Model (ChatGPT)

Quantum Software Engineering (QSE) has recently emerged as a potential research and development direction frequently practiced by many tech joints. However, quantum developers face challenges in optimizing quantum computing and QSE concepts. Quantum developers use the Stack Overflow (SO) platform to report and discuss quantum-related challenges. Also, quantum practitioners use specialized quantum tags to label quantum-related posts in developers' forums. However, these quantum tags referred to more technical quantum aspects than the developer posts. Therefore, categorizing quantum practitioners' questions based on quantum concepts can help quantum developers better identify frequently occurring challenges to QSE. For this purpose, we conducted qualitative and quantitative studies to classify quantum developers' questions into various frequently occurring quantum-related challenges. We extracted 2829 developers' questions from various Q&A platforms using queries and filters based on quantum-related tags. Next, the developers' posts on the Q&A forums were critically analyzed to identify frequently discussed quantum-related challenges and develop a novel grounded theory. The frequent quantum developer challenges identified by analyzing practitioners' discussions in the Q&A forums include Tooling, Theoretical, Learning, Conceptual, Errors, and API Usage. Moreover, using content analysis and grounded theory, the developers' discussions were annotated with commonly reported quantum challenges to develop a ground truth and a novel dataset. A Large Language model (ChatGPT) was used to validate the human annotation and overcome disagreements. Finally, various fine-tuned Deep and Machine learning (D&ML) classifiers automatically classify developer discussions into commonly reported quantum challenges. Additionally, to improve the classification results, we utilized textual data augmentation approaches, such as random deletion, swapping, and insertion with the D&ML classifiers. We obtained average accuracies of 89%, 86%, 84%, 84%, and 80% with FNN, CNN, LSTM, GRU, and RNN classifiers, respectively. This helps quantum researchers and vendors propose solutions and tools to frequently occurring issues for quantum developers.