Communication Dans Un Congrès Année : 2024

Exploring the abilities and limits of generative AIs on specialized translation: a case study on United Nations documents on violence against women

Résumé

Large language models (LLMs) have reshaped the ways in which linguistic tasks are addressed, and the ways in which linguistic-related topics are taught. Their multilingual abilities make them very powerful in the field of machine translation (MT), enabling one to quickly obtain good-quality translations and to easily perform reformulations of the outputs. Although concerns have been raised about the possibility of such systems replacing professional translators, the current state of affairs rather suggests that LLMs still require human post-edition in most cases. It is therefore of utmost importance of train students in MT to use LLMs at the best of their ability, and that implies making them aware of their limits, especially in the context of specialised translation where neural MT systems have already shown weaknesses. The present paper aims to explore the abilities of several models such as ChatGPT, Bard, LLaMA, and Mixtral on translating United Nations’ documents on the topic of violence against women between English and French. The interest of such a case study lies on the quasi-systematic availability of UN documents in the six official languages of the organisation, which provides a large amount of data for comparison, and on the diplomatic constraints of the UN that shape its writing style and its lexis. Furthermore, the topic of violence against women calls for specialised knowledge in a field where concepts and terminology may rapidly evolve, which may be challenging for LLMs whose training datasets stop at a specific point of time. In the context of teaching, we argue that such a study illustrates the benefits and the limits of LLMs in MT.

Presentation_ESSE-8.pdf (5.55 Mo) Télécharger le fichier
Origine Fichiers produits par l'(les) auteur(s)
Licence

Dates et versions

hal-04894315 , version 1 (27-01-2025)

Identifiants

  • HAL Id : hal-04894315 , version 1

Citer

Célia Atzeni, William Babonnaud. Exploring the abilities and limits of generative AIs on specialized translation: a case study on United Nations documents on violence against women. ESSE Conference, Université de Lausanne, Aug 2024, Lausanne, Switzerland. ⟨hal-04894315⟩
120 Consultations
89 Téléchargements

Partager

  • More