MRTCNN: A Lightweight Approach for Predicting Metamorphic Relations

Metamorphic testing is a software testing technique that provides both a test case generation strategy and a test result verification mechanism. Its foundation is a set of metamorphic relations, which are basically the necessary properties of the software under test, represented in the form of relationships among multiple inputs and corresponding expected outputs. As the core element of metamorphic testing, metamorphic relations have attracted lots of research interests from different perspectives, among which one major direction is to identify metamorphic relations suitable for certain types of programs. In order to reduce the manual work in the identification process, machine learning techniques have been leveraged to predict valid metamorphic relations for scientific software. In this paper, we present a new approach for predicting metamorphic relations based on the deep learning of the program documentation. In particular, we make use of the text convolutional neural networks in the prediction and validation of proper metamorphic relations. Empirical studies have also been conducted to evaluate the applicability and performance of our approach. The experimental results demonstrate its effectiveness in predicting appropriate metamorphic relations for the testing of various Java programs. Compared with the existing baseline techniques, our approach improves the precision and accuracy of the metamorphic relation prediction process. This study also reveals potential research opportunities for advancing the performance of metamorphic testing.