La gestion des connaissances à l’ère de l’IA
Talk, CNES : Séminaire COMET Knowledge Management, Toulouse
The IRIT attended the COMET workshop on 16 June 2026, where they presented Extracting specialized semantic relations from texts and tables with generative AI models. Abstract: Relation extraction is essential for enriching knowledge graphs, yet specialized domains raise specific challenges such as domain terminology and meaning spread across texts, tables and figures. In this talk, we first present a benchmark of LLMs for relation extraction on the CORE economic dataset, showing that Llama3 consistently outperforms other models and that PEFT-based fine-tuning raises F1 from 0.69 to 0.80, with 50–70% of the training data being sufficient. We then introduce an original task of joint text-table relation extraction, together with the manually annotated ReTaT corpus (Business, Female Celebrities and Telecom domains) linked to Wikidata, which served as the basis for the TRIPLET 2026 challenge at ESWC. A complementarity study demonstrates that joint extraction yields 70% new triples that cannot be obtained from texts or tables alone. Finally, we describe an optimized hybrid pipeline combining LLM-based preprocessing, NER and entity linking ensembles, TARTE table embeddings and a LightGBM classifier, which ranked 2nd in the TRIPLET challenge. These results confirm that naive LLM prompting yields mediocre results, whereas combining classical approaches, table semantics and knowledge graph entity types substantially improves specialized relation extraction.