Error Analysis of a Machine-Translated Poem

Towards a Teaching Guide

Abstract

Abstract
The article is a suggested three-hour practical class on teaching error analysis in a machine-translated
poem in the French to English direction. The poem Femmes de France by Léopold Sédar Senghor was
used in the step-by-step error analysis teaching. The article shows how four machine translation (MT)
systems (Amazon, DeepL, Google and Microsoft) output a typical poem by an African author and how
the analysis of the errors from the four machine translations can be taught to students of the literary MT
class. The poem was first analysed for errors against the original work, which is in French, and then
analysed again for errors against Melvin Dixon‟s English translation (human translation). An adaptation
of Kopenen‟s (2010) error classification model for MT was chosen for this demonstration because the
model was found more suitable for analysing meanings and intentions in the literary genre of poetry than
the other available models. In the end, it is revealed that the error class of mistranslated concept, after
analysis against the human translation (HT), had a more significant number of errors than the same class
of mistranslated concept when analysed in the source poem. This result shows the students that the human
translator is a more critical reference than the poem‟s author when evaluating the MT of poetry