*This article is a part of the series “Artificial Intelligence and Its Applications: Perspectives from Across Kent State,” highlighting the applications of AI in different fields and including insights from students and faculty. Stay tuned for a future article covering design, or check out previous articles about law enforcement, manufacturing or healthcare.
Geoff Koby, Ph.D., professor of Translation Studies and German, certified freelance translator and president of the American Translators Association (ATA), said artificial intelligence is inescapable in translation and interpretation now. In his role at the ATA, the largest professional association for practicing translators and interpreters in the world, Koby has had significant exposure to machine translation.
“Anybody graduating now as a translator is going to be faced with a very different landscape than I encountered 30 years ago when I started translating,” Koby said.
Origins and Evolution of Machine Translation
According to Koby, machine translation traces its roots back to the ‘50s and ‘60s during the Cold War when the United States government funded research on it, but it fell short due to low computing power. In the ‘90s, statistical machine translation evolved, which was a vast improvement, but still not as good as professional human translation.
In 2016, researchers made a big breakthrough with the creation of neural machine translation (NMT). NMT uses artificial neural networks to do word associations, which works similar to predictive typing by ‘guessing’ what comes next. Since the advent of NMT, completing non-complex tasks, such as translating a quick email, can generally be done successfully.
Strengths and Weaknesses of AI in Translation
“For many things, whether you’re going through NMT or AI large language models, the output is good enough for a lot of everyday purposes,” Koby said.
However, for more precise translation such as medical or patent translation, human oversight is essential.
“From my perspective as president of the ATA, our message to the world really is there has to be a human in charge,” Koby said. “And not just any human, but a trained professional translator because only a trained professional can recognize where the machine has gone wrong.”
Koby said that if one critical word is translated wrong, the consequences could be drastic. He points to the "poisoned cookie" analogy to demonstrate this idea.
The analogy suggests that if a human bakes 100 cookies, they will vary in size, shape and texture, but they will all be edible. If a machine bakes 100 cookies, they will be perfectly baked and round, but 1 out of 100 is poisonous. In other words, human translation might be flawed, but it is basically all usable. Machine translation appears unflawed on the surface, but it can make dire mistakes that no human would.
“If you don’t know which one of those 100 or 1,000 words is wrong, you don’t know whether that’s a minor mistake that nobody would care about or if it’s poisoning somebody,” Koby said.
Machine interpretation also faces several issues, such as whether the AI correctly perceives audio and if it can accurately interpret variations such as dialects, slang and mumbling. A human interpreter in charge can actively monitor it and intervene when necessary.
Koby has graded ATA translation tests and part of his research focuses on the assessment and evaluation of translation. He said that the standards of assessment are the same for human translators and machines, but the patterns of error are different.
The assessment process involves two factors: accuracy and fluency. Accuracy is how well the facts and the message translate. Fluency is how well the text is written, including elements like spelling and punctuation.
Koby said that machine translation excels in fluency but often struggles with accuracy. Human translators are aware of nuances and cultural considerations such as specialized terminology, style and locale, while machines are not.
According to Koby, the quality of AI translation also varies by language and corpus. French, Italian, German and Spanish translation from English and into English are the most common forms of translation. These languages have large amounts of text on the internet and the AI translation can come out well, but it is often not congruent. For example, certain laws or technical areas may not be found in other languages and specific terms may be missing.
Another consideration is that languages that do not have a written form cannot be machine translated, which also makes them more difficult to interpret.
Koby said that in the future, professional translators may be shoehorned exclusively into the role of post-editors who spend all their time editing, revising and verifying the machine’s output.
“It’s a double-edged sword, it really is,” Koby said. “On the one hand you look at the technology and say that is really pretty amazing. On the other hand, in my position as president of a professional association of translators, if it gets so good that there’s no need for translators, then our profession goes away.”
Incorporation at Kent State and Future Outlook
Koby’s undergraduate translation courses do not utilize AI in the interest of developing the skills of the translation students, but that doesn’t mean discussion of the topic is absent.
Michael Carl, Ph.D., Modern and Classical Language Studies professor and director of the Center for Research and Innovation in Translation and Translation Technology, has 25 years of experience with machine translation. He is looking for ways to incorporate questions about the future of the field into his Ph.D. seminars.
In the seminar course, Philosophical Foundations of Translation Studies, students will investigate questions stemming from the recent rapid advancement of NMTs, such as how they differ from humans and how they truly operate.
“[It seems] everybody is biased toward assigning agency or some kind of understanding to those machines,” Carl said. “[They say] otherwise we cannot understand why this machine produces such fluent output if it works so differently from how we work.”
One of the major differences between human translators and NMTs is the presence of emotions. Carl is interested in understanding the relationship between emotional attitude and the translation output or outcome. He is also curious to see if machines could ever experience feelings and how that would manifest.
“I’m most excited about this idea from my research—seeing what the relationship between machines and humans is and how humans are different from machines,” Carl said.
While Koby recognizes the threat to the profession, the main point of emphasis for him and the ATA remains the same.
“Our takeaway is that always the professional human translator or interpreter is in charge,” Koby said. “That’s the bottom line.”
This series of articles only scratches the surface of the applications of AI -- we’d love to hear from you! Fill out the following survey to share how you use AI in your work:
###