The Limitations of Neural Machine Translations into / from Hungarian. A Case Study

Abstract

Neural Machine Translation (NMT) systems face challenges when translating to and from Hungarian due to the language’s typological and morphological complexity. Hungarian is an agglutinative language with extensive inflection, rich case marking, and flexible word order, which often results difficulties in accurate morphological segmentation. NMT models frequently struggle to preserve grammatical relations, especially when translating between Hungarian and morphologically poorer languages such as English. Problems commonly arise in handling long suffix chains, verbal prefixes, and agreement features, leading to errors in tense, definiteness, and argument structure. Additionally, free word order and discourse-driven focus constructions complicate alignment and sentence-level coherence in translation. Limited availability of high-quality parallel corpora for Hungarian further constrains model performance, particularly in domain-specific contexts. These factors collectively reduce translation fluency and adequacy. Addressing these issues requires improved morphological modeling and the incorporation of linguistic knowledge to enhance NMT quality for Hungarian language pairs.