Requirements Elicitation for Machine Learning Applications: A Research Preview

The development of software systems is preceded by an important first phase, requirements elicitation, wherein developers establish the intended functionality of a system to be developed in a series of interviews with a customer. These requirements can often be raw, requiring refinement in an iterative process. Traditional software requirements are of the form 'The system shall … ‘, where each requirement is written to be clear, concise, consistent, complete, testable and traceable through development. Because Machine Learning (ML) is stochastic, or uncertain, in the face of unseen data, ML behavior cannot be precisely defined. As such, it is posited that software requirements specific to ML must delineate a window of acceptable behavior, e.g., a plus/minus value for evaluated model metrics. The provision of such is essential for the quality of future ML software systems, as the current trend of large-scale generative ML has brought about a paradigm of fine-tuning, e.g., of Large Language Models, in which pre-trained large models are taken “off the shelf” and made fit for very specific purposes. These specific purposes vary with different use cases, but they are effectively undeclared requirements. The present paper discusses requirements elicitation in the context of ML by considering the four-software engineering requirement elicitation techniques: conversational, observational, analytic and synthetic, providing future researchers with insight for eliciting quality ML requirements.