Comparing Human and Specialized Neural Machine Translation for Accuracy and Efficiency

Abstract

Objectives: The advancement of machine translation has sparked significant interest in research on human-machine translation. This study investigates the effectiveness of neural machine translation (NMT) for various types of text categories. Methods: To test the hypothetical role of NMT, we conducted a comparative experiment on four types of texts: medical, economic, legal, scientific and technical texts. Each text was translated from English into French and Arabic, and the English version was chosen as it is often directly translated into other languages. An expert human translator produced the French and Arabic versions, and to assess the accuracy of the machine translation, we translated the French version into Arabic using MT. The three versions - English, French, and Arabic - were then automatically aligned and distributed to different groups of students to identify linguistic errors in the Arabic version of the MT. The students also ranked the MT translation from closest to the expert's translation to least close. Results: The findings revealed that the machine translation was most effective for economic texts, followed by medical and scientific-technical texts, then the legal texts based on the number of errors.Conclusion: Although the MT successfully conveyed the main ideas, syntactical structures were the main issue. Therefore, improving MT depends primarily on enhancing syntactical structures.