Are NMT and LLMs effective in the translation of dialogues?

Abstract

 AI has made huge strides forward and is increasingly applied in language learning and audiovisual translation. This paper explores the quality of the machine- and LLMs-driven translation of excerpts of naturally occurring conversations scrutinised by scholars. The dialogues are sourced from literature analyses and examples. After outlining the traits of spoken language, this paper delves into an examination of the renderings of the transcripts of spoken interactions from Italian into American English. To do so, the DeepL NMT platform and the ChatGPT and Gemini chatbots are taken into account. The paper findings bring to the fore three main aspects: 1) NMT is qualitatively less accurate and reliable than LLMs-generated translations; 2) human intervention in LLMS-driven output is necessary because of sporadic inconsistencies and mistranslations, and 3) LLMs prompts must be written carefully and mindfully in order to obtain consistent and accurate results, as unspecific prompts may give rise to a less satisfactory output.