Identifying AI Features Within a Software Requirement Specification Using a Cascaded Hybrid Architecture

Abstract

In recent years, there has been an increasing interest in using Artificial Intelligence (AI) capabilities in software engineering practices. The existing Requirements Engineering (RE) verification approaches are still largely designed to address deterministic software behaviors. As such, AI requirements are often difficult to identify within hybrid Software Requirements Specification (SRS) documents. We selected and evaluated three models: TF-IDF with Logistic Regression, BERT-base, and RoBERTa-base. Among the three evaluated models, RoBERTa-base produced the best results. This study presents a semi-automated approach based on a cascaded hybrid architecture. The architecture combines IEEE-guided rule-based preprocessing, a two-stage RoBERTa classification pipeline, and a Human-in-the-Loop (HITL) validation layer to support a trustworthy decision-making process in identifying AI requirements within an SRS document. This led to the development of the AI requirement checker tool. For the evaluation, we used four SRS documents containing 164 extracted requirements, including 11 AI requirements. The results showed an overall accuracy of 95.7%, a recall of 81.8%, and an F1 score of 72.0% for AI requirements identification. The tool also produced no false positives when evaluated on a fully deterministic SRS document. Usability inspection was based on ISO 9241-110, which showed full conformance for task suitability and controllability. The results suggest that the tool can reduce manual review effort while allowing experts to retain control during requirement verification. For the future, the tool would focus on full automation, dataset expansion, and improving performance on larger and less structured SRS documents.