The method, which first appeared in 2014, is based on networks that are modelled on the neuronal structure of the human brain. The special feature of the method can be most clearly seen in comparison to statistical machine translation (SMT): In SMT, large quantities of translated texts are compared before the actual translation and the source and target texts are assigned to each other based on frequency and transferred to a virtual dictionary with grammar rules. These word pairs form the basis of statistical machine translation.
NMT also analyses bilingual texts, with the difference that, thanks to an artificial neural network, grammatical correlations between the source and target language are implicitly recorded. In this way, word dependencies can be recognised and correctly reproduced, especially in longer sentences. The readability of neuronal machine-translated texts is already apparently better than that of statistical texts.
Imagen Gerd Altmann auf Pixabay
