Automated Ethical Requirements Classification in AI: The ERC4AI Tool

Context: This work explores the intersection between Requirements Engineering (RE), Artificial Intelligence (AI), and Natural Language Processing (NLP), focusing on the automatic classification of ethical requirements in AI systems. Objective: To create a classifier model to identify ethical requirements in AI, to help developers integrate ethical considerations from the early stages of projects. Method: We adopted the Cross Industry Standard Process for Data Mining (CRISP-DM) to structure our research. The process began with a literature review, followed by text collection, analysis, and preprocessing. Subsequently, we performed a comparative benchmark using different classifiers based on the XLM-RoBERTa, BERT, and DistilBERT algorithms, applying them to the EthicalRequirements4AI dataset to evaluate the classification performance. Finally, we conducted a case study with data different from those used in training and initial evaluation, to validate the generalizability and applicability of the models in a practical scenario. Results: The binary model with XLM-RoBERTa achieved an F1-Score of 0.90 in the classification of ethical requirements. Conclusion: The research resulted in the creation of the Ethical Requirements Classifier for AI (ERC4AI) and EthicalRequirements4AI dataset, which represent a significant advance in RE for AI, filling this important gap in the literature, essential for risk mitigation in the development of AI systems.}