Neural Machine Translation

The work is devoted to the comparison of neural machine translation systems based on different architectures of recurrent neural network and methods of word representation, in particular from a linguistic point of view. At the first stage, three neural translation systems based on three architectures (I. Sutzkever, K. Cho and using the attention mechanism) were built, and a classification of mistakes was proposed. At the second stage, two machine translation systems were built using different types of word representation (context-based and dependency-based) and their results were compared. During the linguistic analysis, five types of mistakes occurring at the lexical and syntactic level and associated with both syntagmatic and paradigmatic relations between words were identified and explained: missed words, extra words, changed word form, word substitution, the wrong word order. Although the quality of translations in the second stage was low, the comparison showed that although both machine translation systems strive to preserve the semantic similarity of sentences, the system using syntax-based embeddings partially retained the elements of sentence structure and some grammatical forms.