CIAD System for Geographical Entity Detection at TextMine'24
Résumé
This paper outlines CIAD's approach to Named Entity Recognition (NER) in a corpus of nautical instruction. Employing a conventional entity detection approach, our system tackles the NER task through token classification. We handled this classification task using two approaches: 1. NER is performed using different BERT models (BERT-base, Tiny-BERT, and CamemBERT); 2. We also used a GCN-based token classification where a graph is built connecting words in the same context. Our results suggest that within these token classification setups, our models are bounded to an accuracy of around 92% on the test data regardless of the complexity of the used BERT model.
Domaines
Intelligence artificielle [cs.AI]
Origine : Fichiers produits par l'(les) auteur(s)