Introducing FELA - Flexible Entity Linking Approach

With the rapid expansion of digital data, effective mechanisms for transforming raw information into structured

knowledge are increasingly essential. End-to-end entity linking presents a promising solution by disambiguating

entity mentions and aligning them with knowledge bases. However, most existing approaches are tailored to a

single KB, limiting their adaptability and scalability across diverse knowledge resources. To address this limitation,

we introduce FELA - Flexible Entity Linking Approach - a framework designed for seamless entity linking across

multiple knowledge bases. FELA leverages fine-tuned Large Language Models, a generic embedding model, and a

Large Language Model-based reranking module to enhance entity disambiguation. Our approach achieves state-

of-the-art performance on Wikidata entity linking benchmarks, demonstrating its effectiveness and flexibility.

Furthermore, we illustrate FELA’s extensibility by applying it to Agrovoc, showcasing its capability to generalize

beyond Wikidata. This work contributes to the development of more flexible, scalable, and domain-agnostic

entity linking solutions, facilitating knowledge extraction across heterogeneous data sources.