RADAR: Risk-Aware Ambiguity Detection for Regulatory Requirements

Abstract

Ambiguity in natural-language requirements remains a persistent challenge in safety-critical systems, where unclear specifications may propagate into verification failures, redesign effort, certification delays, and downstream safety concerns. Existing ambiguity detection approaches increasingly employ rule-based techniques and transformer-based models; however, many formulate ambiguity analysis as a prediction task using fixed thresholds and symmetric evaluation assumptions. This paper presents RADAR (Risk-Aware Decision Framework for Ambiguity in Requirements), a hybrid framework that combines interpretable linguistic indicators with transformer-based semantic representations through probabilistic fusion. RADAR incorporates a Bayesian cost-sensitive decision layer to adapt ambiguity thresholds under varying risk conditions, enabling explicit control over false-negative and false-positive trade-offs. The framework also incorporates lightweight regulatory grounding and LLM-assisted recommendation support for complianceaware refinement. Evaluation on a corpus of 3,312 regulatory and safety-critical requirements shows that RADAR improves ambiguity coverage (Recall $=0.9022)$. Under higher false-negative cost settings, recall increased to 0.998 without retraining, demonstrating adaptive decision behavior according to operational priorities. These findings suggest that risk-sensitive ambiguity analysis can support practical decision-making within safety-critical requirements engineering workflows.