Self-Admitted Technical Debt classification using LSTM neural network

Abstract

Context: In software development, new functionalities and bug fixes are required to ensure a better user experience. Sometimes developers need to implement quick changes to meet deadlines rather than a better solution that would take longer. These easy choices, known as Technical Debts, can cause long-term negative impacts because they can bring extra effort to the team in the future. One way to detect technical debts is through source code comments. Developers often insert comments in which they admit that there is a need to improve that part of the code later. This is known as Self-Admitted Technical Debt (SATD). Objective: Evaluate a Long short-term memory (LSTM) neural network model to identify design and requirement SATDs from comments in source code. Method: We performed a controlled experiment to evaluate the quality of the model compared with two language models from literature in a labeled dataset. Results: Our model results outperformed the other models in precision, improving average precision in approximately 8% compared to auto-sklearn and 19% compared to maximum entropy approach, however, the LSTM model achieved worse results in recall and f-measure. Conclusion: We found that the LSTM model can classify with better precision but needs a larger training, so it can improve on the detection of negative cases.