stEELlm: An LLM for Generating Semantic Annotations of Tabular Data

The capabilities of LLMs represent a pivotal step in transforming how we manage and interact with information and data. We witness an increasingly pervasive use of such models in various computational tasks. In some preliminary works, attempts to integrate Knowledge Graphs and Large Language Models (LLMs) can be identified, in particular, to perform the classic tasks related to the construction of Knowledge Graphs through semantic annotation of texts. Nowadays, tables are widely used and play a crucial role in creating, organising, and sharing information that could be used to produce factual knowledge to be integrated into a Knowledge Graph. However, table-to-KG techniques through LLM have not been extensively investigated. This paper presents stEELlm, an innovative Semantic Table Interpretation approach obtained by fine-tuning the Mixtral 8x7B model. Conducted experiments demonstrate the capabilities of our model to successfully create semantic annotations of heterogeneous datasets, a scenario where classic approaches based on heuristics tend to fail.