Stylistic Features and Creativity in Machine Translation of Literary Texts

Abstract

Computer-assisted literary translation has become a trending top­ic of discussion with both its possible benefits and negative impact on cre­ativ­ity and voice. We investigate distinctive stylistic features and the degree of creativity in the outputs generated by three English–Turkish machine trans­lation (MT) models: (1) a customized MT model trained with literary texts, (2) a pre-trained MT model with general texts, and (3) an online MT mod­el, namely Google Translate. We focus on two sub-genres within the lit­er­ary domain: fiction and nonfiction. Our analysis of style and creativity is based on human evaluation of samples and on qualitative corpus analysis of full texts, involving identification of creative translation strategies and stylistic fea­tures. We compare the outputs of the three models with translations by two renowned translators, Nihal Yeğinobalı and Belkıs Dişbudak, on the test set. Our investigations of style and creativity involve different methods, yet we consider the possible relationship between the two in light of our find­ings. A higher level of creativity is observed in the human translation and in the fine-tuned model for the nonfiction; and a customized MT model trained with Turkish literary translations generates outputs stylistically closer to the human translation than a pre-trained model or an online MT tool.