Human-LLM Collaboration for Context-Dependent Requirements Engineering Tasks

Abstract

Many requirements engineering (RE) tasks, such as requirements elicitation, documentation, and quality assessment, are inherently context-sensitive: what counts as a missing, wellwritten, or defective requirement varies by stakeholder-intent, domain, process, as well as countless other potential factors. Existing automation, whether rule-based, ML-based, or based on prompting large language models (LLMs), encodes a onesize-fits-all assumption that does not accommodate this contextdependence. The goal of this dissertation is to investigate HumanLLM Collaboration (HLC) as a general paradigm for contextsensitive RE tasks. In HLC, an LLM produces predictions or proposals together with explanations, users validate the output, and the resulting history of accepted or rejected output adaptively improves the accuracy and usefulness of future suggestions. For this research, the primary application area of HLC will be requirements quality assessment. Initial work demonstrates that HLC predicts the defectiveness of the weak word smell more effectively than standard few-shot prompting or fine-tuning techniques. The planned research characterizes how context shapes quality, extends HLC to different RE tasks, and evaluates it with practitioners.