Enriching RDF Data with LLM Based Named Entity Recognition and Linking on Embedded Natural Language Annotations

In this paper, we present a processing pipeline for transforming natural language annotations in RDF graphs into machine-readable and interoperable semantic annotations. The pipeline uses Named Entity Recognition (NER) and Entity Linking (EL) techniques based on a foundational Large Language Model (LLM), combined with a Knowledge Graph (KG) based knowledge injection approach for entity disambiguation and self-verification. Through a running example in the paper, we demonstrate that the pipeline can increase the number of semantic annotations in an RDF graph derived from information contained in natural language annotations. The evaluation of the proposed pipeline shows that the LLM-based NER approach produces results comparable to those of fine-tuned NER models. Furthermore, we show that the pipeline using a chain-of-thought prompting style with factual information retrieved via link traversal from an external KG achieves better entity disambiguation and linking than both a pipeline without chain-of-thought prompting and an approach relying only on information within the LLM.