Contributors: Université Grenoble Alpes (UGA); Laboratoire d'Informatique de Grenoble (LIG); Centre National de la Recherche Scientifique (CNRS)-Université Grenoble Alpes (UGA)-Institut polytechnique de Grenoble - Grenoble Institute of Technology (Grenoble INP ); Groupe d’Étude en Traduction Automatique/Traitement Automatisé des Langues et de la Parole (GETALP); Université Grenoble Alpes (UGA)-Centre National de la Recherche Scientifique (CNRS)-Université Grenoble Alpes (UGA)-Institut polytechnique de Grenoble - Grenoble Institute of Technology (Grenoble INP ); Lingua Custodia; Institut des Langues et Cultures d'Europe, Amérique, Afrique, Asie et Australie (ILCEA4); Laboratoire de Linguistique Formelle (LLF - UMR7110); Centre National de la Recherche Scientifique (CNRS)-Université Paris Cité (UPCité); Université Paris Cité (UPCité); Centre de Linguistique Inter-langues, de Lexicologie, de Linguistique Anglaise et de Corpus (CLILLAC-ARP (URP_3967)); ANR-19-P3IA-0003,MIAI,MIAI @ Grenoble Alpes(2019); ANR-21-CE23-0021,CREMA,Résolution de coréférence pour la traduction automatique(2021)
Abstract: International audience ; Model explainability has recently become an active research field.Many works are published supporting or criticizing attention weights as model explanation. In this work we adhere to the former and analyze attention as explanation for Context-Aware Neural Machine Translation (CA-NMT). Since its evaluation often concerns the evaluation of models in resolving discourse phenomena ambiguity, we perform analyses and evaluations over coreference links in a parallel corpus. We propose a human evaluation over heatmaps, strengthened by a quantitative evaluation based on attention weights over coreference links and with different metrics purposely designed for this work. Such metrics provide a more explicit evaluation of the CA-NMT models than evaluations using contrastive test suites.
Processing Request
No Comments.