It may not be long before automated translation programs catch up with human translators, as machines are evolving with the help of artificial intelligence to “learn” how to translate by mimicking the workings of the human brain.

Major tech companies including Google and South Korean portal website operator Naver have recently rolled out translation software relying on a new AI-based computing system called “neural machine translation.”

Neural machine translation is a new translation algorithm that is distinctive from the widely used phrase-based machine translation system, also known as statistical machine translation. 

While phrase-based methods translate individual words and phrases independently within a sentence and rearrange them -- often leading to awkward mistranslations -- NMT views the entire sentence as one unit to be translated at once.

NMT simultaneously considers different ways to translate a given sentence, and arrives at the most appropriate outcome based on context, which it learns to discern by studying a vast database of translated materials. The system mimics the workings of the human brain and the way it processes and learns information.

Two forerunners in making use of this artificially intelligent computing algorithm include Google’s revamped Google Translate and Naver’s newly released translation app Papago.

Google Translate & Naver’s Papago

Naver rolled out Papago, which comes from the word “parrot” in Esperanto, in August 2016. The app can translate between Korean and English as well as Korean and simplified Chinese using NMT. It also supports translations between Korean and Japanese, although these translations are dependent on statistical methods for now.

Three months later, Google switched its 10-year-old translation system to NMT for eight languages that combined are used by a third of the world’s population -- English, French, German, Spanish, Portuguese, Chinese, Japanese, Korean and Turkish.

Both apps can translate typed text, speech and images of words or sentences into the target language. Both claim to offer heightened translation accuracy compared to phrase-based services, and to be able to improve their translations over time through repeated machine learning.

By adopting NMT, which studies translated materials available on the open web, Google claims it has reduced translation errors by up to 85 percent and tripled its translation computing speed.

On the other hand, Naver says that according to its internal tests, NMT — despite having existed for only one to two years — offers translations twice as accurate as its statistical predecessor, which had been regularly updated for more than 10 years.

When tested by The Korea Herald, the two NMT-powered translators offered generally improved translations between Korean and English compared to their phrase-based predecessors. Yet, both Google Translate and Papago were prone to mistakes when translating longer, more contextually complex or highly idiomatic sentences, signaling room for improvement.

As time passes, these new NMT-powered translation programs will in theory grow smarter, as they study a growing database of translated materials through deep learning — or recurring layers of machine learning.

It’s similar to how Google DeepMind’s AI program AlphaGo learned and mastered the complex game of Go by analyzing and studying an enormous database of professional Go matches.

Google said its eventual goal is to provide highly accurate translations of English into all of the world’s languages, in turn facilitating improved communication between countries.

Naver eyes lead in Korean translations

Though the US tech giant has been a forerunner in AI tech development and has an enormous abundance of global services, Naver is confident in its ability to surpass the US tech giant in offering quality translations involving Korean, its home language.

“Though reviews may differ, Papago has been shown to be stronger at translating colloquial Korean sentences to English and other languages (compared to Google Translate),” said a Naver official.

For example, Papago is able to correctly identify the differing interpretations of the Korean word “bam” — which can mean “night” as well as “chestnut” in Korean, depending on the context of the sentence.

In translating the Korean sentence “Should we roast chestnuts (bam) tonight (bam)?” Papago offered the correct translation “Shall we eat chestnuts tonight?” whereas Google missed the double meaning of “bam” and arrived at, “Do you want to have dinner tonight?”

According to Naver, Papago is consistently learning from the portal website’s rich database of Korean and foreign languages -- including its web dictionaries, professionally translated text on its video-sharing platform V app, webtoons that are translated into multiple languages and information posted on shopping websites supported by Naver.

Looking ahead, Naver plans to expand the languages that Papago supports to include not only Korean, English, simplified Chinese and Japanese, but also Spanish, French, Indonesia, Thai and traditional Chinese.

The Korean internet giant plans to incrementally adopt NMT for these new languages, as it builds up a larger database of translations for these major Asian languages, it said.

Future of translation

As AI-powered machine translators continue to improve, we are all left with the big question -- “Are we nearing an age where learning languages will become obsolete?”

According to Mike Schuster, a senior research scientist at Google’s AI development unit Google Brain, automated translators cannot undermine or replace the value of learning languages.

“Google Translate is a little step toward getting better translations. But there’s a lot more to do,” said Schuster, who recently spoke to reporters in Seoul via a web conference.

“It’s not only the language but the culture we need to understand to make communication perfect,” he said, underscoring the beneficial effects that language learning can bring to a person.

By Sohn Ji-young (jys@heraldcorp.com)