Improving Training Data Quality for Automated Test Case Generation Using Machine Learning and Its Evaluation

We aim to enhance the accuracy of automatic structuring of requirement specifications and automated test case generation by leveraging machine learning techniques. Training with large datasets often faces challenges such as overfitting, extended training time, and high computational cost. Thus, improving the quality of training data is a critical issue. In this study, we propose a training data generation method that can achieve accuracy comparable to that obtained from large-scale datasets, even when only a limited amount of training data is available. Our approach incorporates structured requirement specification documents, similar to test data requirement specifications, into the training set, and applies four different vectorization procedures. The effectiveness of the proposed method is validated through automated test case generation experiments. We quantitatively evaluate the quality of training data produced by the four procedures using two least-squares-based approximations: quadratic and logarithmic. The results demonstrate that selecting requirement specifications for training data with Latent Dirichlet Allocation (LDA) yields the highest-quality training data. These findings confirm the effectiveness of the proposed method for improving automated test case generation.