Gritiya Rattanakantadilok
This study investigates the efficiency of three Neural Machine Translation (NMT) tools—Google Translate, Microsoft Bing, and Amazon Translate—in rendering culturally specific items (CSIs) at both lexical and sentential levels. Five CSIs, mae ya nang (แม่ย่านาง), lek lai (เหล็กไหล), ruesi (ฤาษี), pha yan (ผ้ายันต์) and palat khik (ปลัดขิก), from a selected Thai commercial, were categorised by Katan’s ‘Triad of Culture’ framework. A synthesized typology based on Dickins’s conceptual grid was employed to analyse the translation procedures of each NMT service. The FAR model proposed by Pedersen is employed to assess the subtitle quality. The selected segment of an audiovisual text from a Thai auto insurance commercial was input into the NMT services to evaluate their ability to enhance translation efficiency and quality. The findings revealed that the technical cultural frame predominantly influenced the NMT services, resulting in the focus on linguistic transfer over interpretation. At the lexical level, the NMT tools employed an exoticising strategy, resulting in the translations that were primarily oriented towards the source culture and the source language. At the sentential level, the performance of NMT systems deteriorated, with several mistranslations identified, though omission errors were not present. The outputs also lacked fluency, necessitating post-editing. Despite literature suggesting improvements in NMT output quality, these NMT engines have yet to significantly impact professional translation for the Thai-English language pair. For low-resource languages such as Thai, the role of the human translator remains indispensable and cannot be fully replicated by current machine translation systems.