Addressing the Reliability and Security of LLMs usage with an Integrated Information Science and NLP Architecture

• Bruno H. Brito UFSCar
• Roney L. de S. Santos UFBA

DOI: https://doi.org/10.5753/wperformance.2026.22012

Resumo

This research addresses organizational data transformation by integrating Information Science (IS) and Natural Language Processing (NLP).We argue that Large Language Model (LLM) failures in data interaction tasks (instantiated via text-to-sql), such as hallucinations, stem from applications detached from information governance. We propose an interdisciplinary architecture employing a Knowledge Graph as an Intelligent Proxy to centralize metadata and ontologies. Validation through nine use cases demonstrates the architecture is functionally effective, generating accurate database queries matching ground-truth and identifying invalid relationships without exposing raw transactional data to the LLM.

Referências

Caseli, H. M. and Nunes, M. G. V. (2024). Processamento de Linguagem Natural: Conceitos, Técnicas e Aplicações em Português. BPLN, 2 edition.

DAMA International (2017). DAMA-DMBOK: Data Management Body of Knowledge. Technics Publications, 2 edition.

Davenport, T. H. (1997). Information Ecology. Oxford University Press.

Evans, E. (2016). Domain-Driven Design. Addison-Wesley.

Google DeepMind (2024). Gemini: A family of highly capable multimodal models. Accessed: 2026-01.

Guizzardi, G. and Guarino, N. (2024). Explanation, semantics, and ontology. Data & Knowledge Engineering.

Minaee, S., Mikolov, T., Nikzad, N., Chenaghlu, M., Socher, R., Amatriain, X., and Gao, J. (2025). Large language models: A survey. arXiv:2402.06196.

Nonaka, I. and Takeuchi, H. (2008). The Knowledge-Creating Company. Harvard Business School Press.