Well-written Knowledge Graphs Most Effective RDF Syntaxes for Triple Linearization in End-to-End Extraction of Relations from Text (Student Abstract)

Celian Ringwald (1, 2) , Fabien Gandon (1, 3) , Catherine Faron (2, 4, 1) , Franck Michel (3, 1) , Hanna Abi Akl (1, 5)

1 WIMMICS - Web-Instrumented Man-Machine Interactions, Communities and Semantics
2 UniCA - Université Côte d'Azur
3 Laboratoire I3S - SPARKS - Scalable and Pervasive softwARe and Knowledge Systems
4 I3S - Laboratoire d'Informatique, Signaux, et Systèmes de Sophia Antipolis
5 DSTI Labs - Data ScienceTech Institute

Large generative language models recently gained attention for solving relation extraction tasks, notably because of their flexibility. This is in contrast to encoder-only models that require the definition of predefined output patterns. There has been little research into the impact of the syntax chosen to represent a graph as a sequence of tokens. Moreover, a few approaches have been proposed to extract ready-to-load knowledge graphs following the RDF standard. In this paper, we consider that a set of RDF triples can be linearized in many ways and we evaluate the combined impact of language model size as well as different RDF syntaxes on the task of relation extraction from Wikipedia abstracts

Domaines: Traitement du texte et du document; Algorithme et structure de données [cs.DS]; Apprentissage [cs.LG]; Web

Celian Ringwald, Fabien Gandon, Catherine Faron, Franck Michel, Hanna Abi Akl. Well-written Knowledge Graphs Most Effective RDF Syntaxes for Triple Linearization in End-to-End Extraction of Relations from Text (Student Abstract). AAAI 24 - 38th Annual AAAI Conference on Artificial Intelligence, Feb 2024, Vancouver, Canada. . ⟨hal-04526132⟩