Evaluating the translation of deontic and epistemic modal expressions in legal texts: a comparison between human translators and the GPT-4 model

The inherent differences between languages and the distinct origins of legal concepts often lead to terminological asymmetries, posing significant challenges for legal translators. The emergence of large language models such as GPT-4 has created new opportunities for legal translation. This study examines how GPT-4 handles modal expressions in legal translation compared to human translators within Hong Kong’s bilingual common law system. Modality is a central linguistic feature of legal discourse. It is instrumental in expressing balanced viewpoints, acknowledging legal uncertainty and delineating rights, permissions, obligations and responsibilities. Its translation poses particular challenges due to its polysemous nature and contextual nuances. Using a Chinese-English parallel corpus of 50 Hong Kong judgments with their human and GPT-4 translations, this study analyzed shifts in modal categories, values and orientations during translation. The findings reveal that while GPT-4 maintains basic modal patterns and generally adheres to legal translation norms, its translations exhibit more modality shifts than those produced by human translators. In addition, GPT-4 shows less diversity in its handling of modal expressions. These findings contribute to understanding AI capabilities in legal translation and offer practical insights for optimizing prompt engineering for AI-assisted legal translation while maintaining translation quality in multilingual legal contexts.