Human–Machine Interaction in Translation: Evaluating Post-Editing Strategies in Neural Machine Translation

Abstract

Right now, there is a big change going on in the business of translation. In the past, no one could stop people from being artistic. These days, however, there are many rules about how people can use computers. The speed of this change was sped up by machine learning and neural machine translation (NMT). Post-editing (PE) is now the main way the company makes money. Even with this change, most of the study we have done so far only sees PE as a way to fix things. It doesn't look at the hard strategy choices that human translators have to make when they work with language files that were made by machines. That gap is filled by this study, which looks at the real-world connections between post-editing methods and a number of outcomes, such as brain work, translation quality, and time savings. We look at 177 translation parts from Technical, Medical, Legal, and General genres to see how different methods, ranging from Minimal Correction to full MT Override, change the connection between what NMT produces and what professionals expect. Our results show that choosing a strategy can help you guess how much cognitive load you will be under, and that "interactional efficiency" is a better way to measure output than standard automated metrics. This study adds a verified interactional model of post-editing, giving important information for improving the design of translation technology and the spread of technology use in the workplace. The study changes the way we think about PE as a way for people and AI systems to work together to learn by mixing Psychometrics with behavioral data. In a world where things are becoming more and more robotic, this makes sure that good translation will still be possible.