In the present master's thesis, we are dealing with the analysis of machine translation performed with Google Translate for the English-Slovenian language pair. Google Translate is built on the principle of artificial neural networks that mimic the structure and process of the human neural system. Neural machine translation systems surpass previous types of machine translation systems in the quality and suitability of translations regardless of the type and content of the text. Using an eye tracking tool, a methodology derived from the field of psychology, we investigated the readability of three stages of machine translations – raw machine translation, post-edited and completely corrected translation. We analysed translations in terms of the impact on the reader, which is the concluding point of the translation process, and we also analysed the impact of different categories of errors on the readability of the translations, which are one of the indicators of translation quality. With error evaluation we wanted to determine whether it is possible to pinpoint the most problematic group of errors, ie. errors that most hinder the readability of the translation. Said errors must be taken into account and resolved, especially in the post-editing process of machine translations and in the construction and machine learning of machine translation systems. The purpose of the paper is to present the differences between the readability of a raw machine translation, a post-edited machine translation and a completely corrected machine translation, and to present the level of impact of individual semantic and grammatical categories errors.
|
