ISOprep: an LLM-driven pipeline for semantics-preserving processing of neutralized requirements according to ISO 29148

ABSTRACT:

AI in Requirements Engineering (RE) relies on industrial data, yet safety and privacy risks limit its use. While the GDPR distinguishes only between anonymization and pseudonymization, we use neutralization as a semantics-preserving technique. In AI-supported RE, data heterogeneity and cross-domain variability impede model training. We propose guidelines for semantics-preserving preprocessing for RE datasets based on ISO 29148 criteria, showing that neutralization does not compromise semantics. The approach enables industry–academia collaboration through AI-assisted RE in product development.