Open Access
Abstract
Designing test cases manually for banking software is still a time-consuming endeavour that frequently finds it hard to match the pace of modern agile development processes and changing regulatory requirements. Recent progress in artificial intelligence research showed some promising results in automation of various phases of the testing life cycle. However, most of such solutions use cloud-based models, require large amounts of data, or generate unstructured output, which makes them challenging to implement in an audit-proof way. In this paper, a framework for automated generation of testing scenarios and test cases based on small language models is presented. The solution comprises of a requirement extraction module (SLM-1), which extracts structured testing attributes from user stories and uses deterministic logic rules to ensure that the extracted information complies with the problem domain. The next stage involves validating the output of the requirement extraction phase, followed by the generation of a test scenario and individual steps of testing. Each generated test case in this framework will be traceable to a specific requirement label from user stories. The experiments conducted showed that the approach was able to deliver high production-level extraction performance and provided a structure and reviewable output. Results indicate that a hybrid approach using small language models, trained on particular domains, combined with deterministic test generation algorithms may serve as a feasible and implementable approach to support AI-based testing in highly regulated software domains. The research provides important insights into building enterprise-scale automated testing workflows for advanced banking applications.
Keywords
Graphical Abstract
