Machine translation errors: Problems of classification

The paper considers the problems of classifying machine translation errors. Its first part reviews some approaches to evaluation of machine translation quality and to classification of errors that machine translation systems tend to make. The other part of the paper describes an original taxonomy of machine translation errors — the targeted one. It has been devised specifically to classify the errors central to translation of connectives (from Russian into French). To date, there have been no such studies for this pair of languages. The proposed classification includes two groups of errors: ($i$) grammatical/lexical errors in the translation of the text chunk where a given connective occurs; and ($ii$) errors in the translation of a connective itself. This study uses a parallel Russian–French corpus that stores Russian source texts and their reference — made by professional humans — translations into French. The corpus totals 300 thousand sentences (about 4 million words). The source texts where connectives occur have been used to generate machine translations by two automated systems.