Guest Post: Can Machine Translation Stand Alone?
March 7, 2019
This post, written by Lee Jun-ho, was originally published in the Korea Times.
As mentioned in a previous commentary, translation is beyond the transfer of superficial meaning, but many arguments still only focus on this and its accuracy.
A machine translation service provider reports that the rapid development of technology will make machine translation more accurate and human translators may lose their jobs. An engineering scientist even warned that people will not pay for a translation service once machine translation accuracy is around 90 percent. However, these arguments are partially flawed in that they do not consider how the machine translation engine works and its impact on how translation is created and used.
First, it is necessary to differentiate day-to-day communication from business communication. Casual communication, such as reading machine translated Twitter feeds or reviews of hospitality services, can accept 10 percent inaccuracy, as long as the core message is delivered.
In the business context, however, 10 percent inaccuracy can pose huge financial and reputational risks. To make matters more complicated, current machine translation uses the neural network, which has been known as the "black box." Therefore, no one can predict exactly in which sentence or which part of a sentence the machine translation will make errors. Even worse, it is extremely difficult to pinpoint which part of a neural network is responsible for the error. Of course, if there is any industry that can tolerate this random high risk, human translators will have no choice but to leave the industry.
So what should be done to mitigate this innate risk of machine translation?
The outcome of machine translation needs to be corrected with well-trained human translators and this new type of translation work process is called MTPE (Machine Translation Post Editing). MTPE is emerging in the industry and many researchers are coming up with new findings about it. MTPE may reduce some of the workload of translators compared with translating from scratch and contribute to higher productivity.
However, enhanced productivity does not mean well-trained translators are not necessary. First, productivity and quality of translation do not go hand-in-hand all the time. Rather, too much focus on productivity can hamper the quality of translation. Second, as no one knows where errors are, experienced human translators have to review the document from beginning to end in a precise manner.
Separate from the translating work itself, machine translation needs human support for its assessment. It is humans who can and should make a judgment call for the 90 percent or 100 percent accuracy anyway. Furthermore, translation quality ― not accuracy ― assessment requires a lot of expertise and experience in translation, and accuracy is only a part of that equation. Therefore, human engagement is crucial for further advancement of machine translation.
One thing to clarify is that this commentary is not all about denying the hard work of machine translation engineers and researchers. Machine translation has made a tremendous leap forward and is much faster than human translation, but it is a fair diagnosis that machine translation is nowhere close to perfection. Hence, the right word to describe the future relationship between human translators and machine translation cannot and should not be "replacement" but "collaboration".
Lee Jun-ho is a professional conference interpreter specialized in information technology. He has been teaching translation at a graduate school and other universities for seven years, completed PhD coursework and studies intensively the feasibility of machine translation.