Cultural Discourse in African Poetry: Output by Human Translators and Machine Translation Systems

The article reports on how two European translators (Etienne Galle and Andr Bordeaux), two General machine translation (GMT) systems (DeepL and Amazon Translate) and a Custom-built Microsoft Azure translator engine render the Culture-Specific Items (CSIs) in two anthologies of an African Author, Wole Soyinka, into French.The study seeks to fill a knowledge gap, namely the question of whether and to what extent existing Neural Machine Translation systems that are trained preponderantly with texts produced in Western contexts take CSIs in texts written by authors from a non-Western cultural background into account.Following Aixel's (1996) model for identifying and categorising CSIs, CSIs were first noted in the corpora's human, Deepl and Amazon Translate French translation versions.A Custom Translation Engine (CTE) was thereafter built and trained on Microsoft Azure with parallel data of about 14,000 English/ French African poetry sentences.CSIs were again noted after the CTE translated the anthologies.25 CSIs were identified in the human and machine French translations of the two anthologies studied; more than two-thirds of this number are proper nouns, and the rest are common expressions.The results showed that most CSIs identified in the translation by CTE were lexically and orthographically similar to those identified in the human translation (HT).The results also indicated that CTE's output of CSIs was less monotonous than that of GMT systems.