Traduction de formules de salutation français–espagnol par l’intelligence artificielle : analyse traductologique et typologie d’erreurs Translating French–Spanish Greeting Formulae through AI: A Translation-Studies Analysis and Error Typology

Abstract

Greeting and closing formulae are interactional routines that play a central role in intercultural communication and raise specific challenges for translation between French and Spanish. Although these units have been widely studied from pragmatic, sociolinguistic and conversational perspectives, their treatment by current AI-based translation systems remains insufficiently explored from a translation-studies approach. Methods: This article analyses a bilingual corpus of 361 occurrences extracted from digitised textual and oral corpora, accessible through databases such as Frantext, CREA, and CORDE. The study first classifies the formulae according to their global position in the interaction, their degree of pragmatic fixation, and their local conversational position. It then compares the translations produced by DeepL Pro, a neural machine translation system, and ChatGPT-4o, a generative artificial intelligence system capable of performing translation tasks. Results: The findings show that both systems encounter difficulties when the pragmatic value of a formula depends on context, register, sociocultural convention, or sequential position. The most recurrent problems include errors of interpretation, such as counter-sense, false sense, and loss, as well as errors of reformulation, including omissions, barbarisms, and improprieties. These errors are particularly visible in polyfunctional formulae, whose translation requires the identification of whether the expression opens, maintains, or closes an interaction. The study demonstrates that translating pragmatic phraseological units requires more than lexical equivalence; it also demands contextual, interactional, and intercultural competence. The proposed error typology contributes to the evaluation of AI-assisted translation and highlights the need to integrate pragmatic and discourse-based data more effectively into neural and generative translation models.