Persisting grammatical errors of machine translation

Abstract

The paper examines the quality of machine translation (MT) and maps its advancements in terms of different approaches. It compares two English-to-Slovak MT outputs generated by Google Translator (GT): one based on the statistical approach (Statistical Machine Translation, SMT) and the other on the neural approach (Neural Machine Translation, NMT). The study evaluates the quality of MT outputs in the context of typologically different languages – English, which is mostly analytic, and Slovak, which is mostly inflectional. It uses a sample of journalistic texts that are frequently translated by machine translators due to their wide range of vocabulary and variety of topics. The research results indicate that NMT, compared to its predecessor SMT, has significantly improved in almost all framework categories. The NMT output is much more fluent, sounding more natural and comprehensible. In contrast, shortcomings can be found in the omission of lexemes, literal translations, or in the lexemes with multiple meanings (regarding polysemous or homonymous words). In such cases, neural MT may struggle to select the appropriate fit-in-context meaning; moreover, these lexemes can further shift the meaning of the entire sentence, clause, or even utterance.