The relationship between linguistic distance and neural machine translation quality

Abstract

Among all the factors that may contribute to the output quality of a translation, linguistic distance between the source language and target language had been largely cast aside. Relatively recent developments in linguistic distance research, away from lexical approaches and toward syntactic approaches, have made it possible to apply linguistic distance more methodically. This thesis aims to answer the question whether the neural machine translation quality drops as translated languages get more linguistically distant. To reach this answer in relation to machine translation, a survey was conducted in which participants were asked to evaluate machine translation outputs from different software and on different texts based on questions relating to different error types. Different participants who spoke both the source language Turkish and also increasingly more distant languages to Turkish at an advanced level were found, in order to capture the effect of a wide spectrum of language distance. The results from a relationship between linguistic distance and machine translation quality provide an experimental background for future research regarding this relatively unexplored relationship by raising specific questions about sensitivity towards linguistic distance in building machine translation tools.