Machine translation for language access in government settings

Recent legislative mandates have expanded language access in government services, yet research related to the integration of machine translation (MT) remains limited. This study evaluates the quality and efficiency of Neural Machine Translation (NMT) systems (DeepL, Google Translate) and Large Language Models (LLMs) (GPT-4) in translating government-based legal documents from English to Spanish. Methodologically, the study involved twenty-seven professional translators who conducted human translation (HT), machine translation post-editing (MTPE), and quality evaluation. Translation quality was measured using an adapted Multidimensional Quality Metrics (MQM) framework, while technical and temporal post-editing efforts were measured via keylogging software. The findings indicate that (a) MTPE significantly reduces translation completion time compared to HT; (b) GPT-4, an LLM, achieves higher overall quality scores than traditional NMT engines, including DeepL and Google Translate; and (c) MTPE and HT perform similarly in overall quality. The study underscores the potential of LLM-based translation technologies, combined with professional human post-editing, as a high-quality and efficient solution to meeting growing language access demands in government contexts. These findings offer critical insights for policymakers and translation professionals in public services.