Artificial Intelligence (AI) presents significant opportunities for new business ventures within the industrial domain. However, its adoption introduces a broad spectrum of risk sources across dimensions such as cybersecurity, reliability, robustness, safety, transparency, or human oversight. Simultaneously, organizations are confronted with an increasing amount of digital and AI-based legislation in the international environment. This leads to a fragmented landscape of legal and risk-related requirements, often inconsistent, overlapping, and duplicated. This results in high uncertainty and bureaucratic burden during the development and deployment of AI products. To address these challenges, a systematic, responsible, and efficient approach is required to handle the complex legal requirements as well as the emerging wide range of risk sources. Responsible AI Engineering describes such a systematic approach, enabling the efficient and innovative development of AI products while ensuring compliance with the complex and fragmented legal landscape. In this paper, we introduce the Generative AI Risk Management (GARM) methodology as an instantiation of Responsible AI Engineering. We demonstrate through a concrete industrial GenAI-based product how the GARM methodology enables organizations to drive AI innovation compliantly and with heightened risk awareness.
Blockchain is an emerging technology that facilitates the elimination of intermediaries or third parties, thereby reducing costs and enhancing efficiency. Over the past decade, blockchain has been integrated across various industries, yet its wide-scale adoption within the financial industry remains uncertain. This paper seeks to aid financial institutions in adopting blockchain technology by developing a maturity model. This model is intended to assist organisations in assessing their current maturity level and charting a way forward towards a more enhanced adoption of the technology. We construct an initial model based on literature on existing maturity models in other fields of study. Subsequently, we refine the model through the Delphi approach which is a qualitative, systematic, and iterative panel method. We conduct the Delphi method in three rounds with experts in the field, yielding a final maturity model and a corresponding questionnaire which allows financial institutions to conduct a self-assessment of their blockchain adoption capabilities. We discuss the validity of the model and provide a proof of concept through a self-assessment exercise by a participating institution. The panel participants and contributing institutions concurred that the resulting maturity model and the corresponding self-assessment questionnaire are relevant and useful tools for the purpose of assessing their current maturity and informing future actions to enhance their blockchain adoption capabilities.
Maintaining consistency between business process models and their textual descriptions is critical for operational clarity, compliance, and communication. However, as process models evolve, updating documentation remains costly and error-prone. Existing methods often require manual rewriting or full text regeneration, discarding valuable domain-specific language. This paper presents an edit-based synchronization approach that incrementally updates textual descriptions to reflect changes in Business Process Model and Notation(BPMN) diagrams while preserving unaffected content. We propose two complementary algorithms: the Longest Common Execution Subsequence (LCES) approach for balanced acyclic models, and a heuristic beam search for more complex structures with loops and unbalanced gateways. The resulting transformation steps are translated into structured prompts guiding a large language model to produce minimal, style-consistent revisions. A prototype system demonstrates high semantic accuracy and stylistic coherence across diverse process evolution scenarios.
We introduce NarrativeMind, a previously-unexplored cooperative neural-symbolic decoder in Arabic NLP that dynamically injects dialect-specific cultural constraints during generation. Our approach achieving +3.5 BLEU improvement over AraBERT (from 26.3 to 29.8) while reducing MSA bias by 58 patterns in (wisdom in words, blessing in currency) and (jinās) wordplay as demonstrated in (the speaker said to the killer). Unlike rigid frameworks, NarrativeMind adapts fluidly across Modern Standard Arabic and six regional dialects, particularly benefiting under-resourced Maghrebi varieties. Our real-time multi-dialect collaboration employs adaptive constraint weighting, optimizing both BLEU coherence and our novel CulturalScore metric. This metric derives from 2,500 expert-annotated templates spanning classical (maqāmāt) to contemporary , ensuring comprehensive cultural representation. MADAR corpus evaluation (n=12,000) demonstrates substantial improvements: BLEU scores reached 29.8 ± 0.4 versus 27.1 ± 0.3 for baselines, with dialectal accuracy achieving κ = 0.76 compared to 0.70. Human evaluation involving 15 linguists and 185 native speakers validates 82.5 p < 0.01 ), confirming effective cross-regional story co-creation while preserving dialectal integrity.
To address the challenges of traditional knowledge graph construction (KGC) methods, this paper proposes a framework AdaOntoKG that combines ontology-driven and adaptive chain-of-thought reasoning (Adaptive CoT). The framework builds and expands domain ontologies with LLMs, and guides triple extraction through Zero-Shot-CoT and Few-Shot-CoT, with ontology constraints ensuring semantic consistency and structural validity. Experiments on multiple benchmarks show that AdaOntoKG outperforms mainstream baselines in F1 and AUC, while achieving high ontology conformance. Overall, AdaOntoKG enhances accuracy, controllability, and domain adaptability, providing a scalable solution for constructing high-quality knowledge graphs.
