Combining language models for knowledge extraction from Italian TEI editions

This study investigates the integration of language models for knowledge extraction (KE) from Italian TEI/XML encoded texts, focusing on Giacomo Leopardi's works. The objective is to create structured, machine-readable knowledge graphs (KGs) from unstructured texts for better exploration and linkage to external resources. The research introduces a methodology that combines large language models (LLMs) with traditional relation extraction (RE) algorithms to overcome the limitations of current models with Italian literary documents. The process adopts a multilingual LLM, that is, ChatGPT, to extract natural language triples from the text. These are then converted into RDF/XML format using the REBEL model, which maps natural language relations to Wikidata properties. A similarity-based filtering mechanism using SBERT is applied to keep semantic consistency. The final RDF graph integrates these filtered triples with document metadata, utilizing established ontologies and controlled vocabularies. The research uses a dataset of 41 TEI/XML files from a semi-diplomatic edition of Leopardi's letters as case study. The proposed KE pipeline significantly outperformed the baseline model, that is, mREBEL, with remarkable improvements in semantic accuracy and consistency. An ablation study demonstrated that combining LLMs with traditional RE models enhances the quality of KGs extracted from complex texts. The resulting KG had fewer, but semantically richer, relations, predominantly related to Leopardi's literary activities and health, highlighting the extracted knowledge's relevance to understanding his life and work.