Why Translation AI Results Are Different
The more I learn English, the more I feel sorry for the results of the translator. I speak English on the phone three times a week, and when I can't think of an expression, I use Papago, which is really convenient, but sometimes it's unnatural. (It's been a while since I've used Google Translate), what do you think? Do you use a translator often?
Introducing the “DeepL” Translator
Have you heard of DeepL, a German company? DeepL started supporting the Korean language last January, emerging as a hot translation service with a lot of reviews saying that it is more natural than Papago. In particular, the ability to identify nuances.
DeepL, an AI service company based in Cologne, Germany was established in 2009 and introduced the service for the first time in 2017, but it did not support Korean from the beginning. It is said that it was rated better in its own blind test to the extent that it called itself 'The world's most accurate translator'.
Anyone can use the DeepL translator for free on the website, and the paid version of DeepL Pro translates quickly and without capacity limitation. The Korean version is scheduled to be launched in August. In addition, it can be distributed in the form of an API (Application Program Interface) and applied to other systems to mount it.
There are more than 1 billion users worldwide and supports a total of 31 languages. Beyond text translation, it also supports file translation such as, ‘pdf/docx/pptx.’ (In fact, this function itself is not special.) The beta service ‘DeepL Write’ is a tool that corrects sentence expressions such as grammar, punctuation, tone, and nuance. Currently, they support English (UK), English (U.S.), and German.
Why DeepL Is More Natural
Like other translators, DeepL uses deep learning based on artificial neural networks, However, artificial neural networks are different. Neural networks are an AI operation that mimics the neural signal system in the human brain, and it is a structure in which nodes corresponding to neurons are connected in several layers to derive optimal results.
Google / Papago: Recurrent Neural Network (RNN - Recurrent Neural Network)
DeepL: Convolutional Neural Network (CNN - Convolutional Neural Network)
RNN is a method that reflects the 'order information' of the data and reads sequentially 'left → right'. Google and Papago are adopting this approach. It specializes in sequential data learning by processing input and output in sequence units.
On the other hand, CNN is a structure that extracts 'features themselves', not the order information of data, and identifies patterns based on them. It is mainly useful for image processing such as object and category recognition in images, and DeepL applied it to text translation. As a result of translating Hyundai Motor's official report, DeepL was working hard on 'translation'.
Korean original text - Connectivity is a key element of MECA, and a high-quality and stable communication network must be supported to enable smooth technology operation.
DeepL Translation - Connectivity is a key element of MECA, and a high-quality, reliable telecommunications network is essential for smooth technology operations.
Papago translation - Connectivity is a key element of MECA, and a high-quality and stable communication network must be supported to enable smooth technology operation.
In order to generate accurate translations without disruption, DeepL states that “it's possible only when the communication network is supported” as 'essential for' and for Papago as 'must be supported to enable'."Consistency is also very important in translation," said DeepL CEO Jaroswaf Kutiwowski, adding, "We are adjusting the message to be accurately delivered according to sentences or paragraphs." For example, DeepL represents Washington as the U.S. government, and translates ‘Braised Short Ribs’ as ‘Braised Short Ribs.’
Korean and Japanese Market Targeting
DeepL, which already shows strength in European languages such as French and German. Nevertheless, why does CEO DeepL expect the Korean market to grow into one of the top five global markets (Germany, the United States, France, Japan, and Korea) within a few years.
"There are not many people who use Korean as their first language around the world, but demand for Korean is rising tremendously," he said, adding that the Korean translation market is bound to grow because the process of translating Korean into English is very complicated.
In fact, the Korean-speaking population is about 80 million, ranking only 23rd among the world's languages. Nevertheless, the reason why AI companies focus on Korean AI translation is because it is "difficult, but demand for translation increases."
It can function as a test stand, and furthermore, there are more and more applications that require accurate translation of K contents. It is said that it is a difficult task because the order of Korean words is different from English and German, and there are so many different types of honorific expressions, formal expressions according to the context, and verb usage.
Google CEO said, "Korean and Japanese services, which are very different from English, are a kind of challenge," and explained why he started Korean and Japanese services for the first time other than English on AI chatbots. In addition, Korea and Japan predicted that the acceptance of AI services would be high as they are countries widely receptive to adopting AI.
An AI Translation Evolution
The evolution of AI translation services like DeepL has made significant strides in refining the accuracy and naturalness of translations, especially when it comes to the Korean language. By using a different AI architecture based on Convolutional Neural Networks (CNN) as opposed to Recurrent Neural Networks (RNN) utilized by Google and Papago, DeepL has managed to provide translations that better capture the nuances of the language.
The market demand for improved translation services, particularly for languages like Korean and Japanese, is on the rise due to their complex nature and the growing interest in K-contents worldwide. However, despite these advancements, AI translations are still a work in progress, and the need for human involvement remains, given the complexity of languages and the intricacies involved in their translation. As AI translation technology continues to evolve, we can expect more accurate, reliable, and context-aware translation services in the future.
