Decoding Figurative Language in Journalism Translation: Human and LLM Translation of English-Vietnamese Metaphors

Abstract

The increasing use of large language models (LLMs) in professional translation has reshaped multilingual news production. While prior research has reported gains in fluency and surface-level accuracy, limited attention has been paid to LLMs’ ability to translate figurative language, which plays a key role in meaning construction and evaluative stance in journalistic discourse. This study investigates the extent to which state-of-the-art LLMs achieve functional and cultural equivalence in the English–Vietnamese translation of journalistic metaphors, in comparison with professional human translators. Adopting a mixed-methods comparative design, the study analyzes a purposively selected corpus of 50 metaphorical sentences from major international news outlets. Translation outputs produced by two LLMs (GPT-4o and Claude 3.5) under two prompting conditions (zero-shot and advanced prompting) were compared with a human benchmark generated by senior Vietnamese news editors. All translations were evaluated through blind expert assessment based on four criteria: semantic accuracy, naturalness, cultural equivalence, and journalistic style. The findings show that advanced prompting significantly improves AI performance, particularly in semantic accuracy and stylistic fluency. However, LLMs continue to underperform in cultural equivalence, often exhibiting literalism and reduced metaphorical resonance. These findings underscore the indispensable role of human expertise as cultural and pragmatic mediators, suggesting that while LLMs accelerate production, human transcreation remains essential to preserving the rhetorical integrity of multilingual journalism.