Abstract In the era of rapidly growing digital data, organizations and researchers increasingly encounter unstructured, noisy, and heterogeneous sources, such as social media posts, sensor logs, and scientific publications. Extracting meaningful and structured knowledge from these sources remains a significant challenge, as traditional rule-based or supervised information extraction methods often fail to handle noisy or incomplete data. This paper proposes a novel hybrid framework that integrates generative AI, specifically large language models (LLMs), with knowledge graph construction techniques to automatically extract, organize, and refine knowledge from unstructured data. The framework incorporates modules for entity and relation extraction, noise filtering, denoising, incremental graph integration, and validation, ensuring the reliability and usability of the constructed knowledge graphs (KGs). To evaluate the proposed approach, we consider two representative use cases: processing social media-style text and ingesting semi-structured scientific abstracts. Comparative analysis against traditional NLP pipelines, LLM-only extraction methods, and other hybrid approaches demonstrates that the framework achieves higher precision and recall, reduces redundancy, and produces more complete and accurate knowledge graphs. The resulting KGs are well-suited for downstream analytics tasks, including graph-based reasoning, information retrieval, and decision support. These results indicate that integrating generative AI with knowledge graph construction provides a robust and scalable approach for transforming messy, unstructured data into structured, actionable knowledge.
Mahamuni et al. (Sat,) studied this question.