The dawn of artificial intelligence translation

The dawn of artificial intelligence translation

Google Translate enables the world to communicate with each other

During the World Cup in Russia this summer, Google Translator's use of traffic burst into a wave, because fans and Russians from all over the world need to use this translation software to communicate with each other.

According to Google's data, the words "stadium" and "beer" were particularly used among all the translated phrases during the World Cup.

In the era when almost everyone has a smart phone, who needs to look up a dictionary or a manual of foreign language phrases, two-thirds of 16-34-year-olds use mobile translation software to solve language problems when traveling abroad.

However, mobile translation software is by no means perfect. One in five people travel abroad because of mobile phone translation errors that cause communication barriers or misunderstandings, especially in non-mainstream machine translation technologies.

In addition to translation problems, computer programs will have technical errors, and some netizens will find that typing "dog" 18 times will cause unintelligible English translation.

The dawn of artificial intelligence translation

Low level error

So why are there "low-level errors" in translation in this era of deep learning of supercomputers and machines?

One of the big problems is that the same word usually has more than one meaning. The same-synonyms not only cause tourists to leak, sometimes even government documents will be wrong.

For example, in July this year, the British government's Brexit white paper mentioned "democratic exercises", which translated into German into "demokratischeÜbung", and the word "Übung" refers to physical exercise, not exercise of rights.

Human level

In order to solve similar problems, translation software continuously optimizes the method of machine learning. The software will use some translated text and then refer to the above to select the most appropriate translation of a word.

Earlier this year, Microsoft announced that the quality of their artificial intelligence translations has reached human level. They will become a group of Chinese news machine-translated into English. An independent group of experts evaluated the AI ​​translation results and placed professional human translators. The translation results are comparable.

Microsoft says that they can make this breakthrough, the most important thing is that they use the help of artificial intelligence deep neural networks and reference large-scale machine translation data.

The dawn of artificial intelligence translation

To put it simply, artificial intelligence translation first generates a "draft". This draft then undergoes a continuous process of text comparison, comparison, learning, and human translation.

Artificial intelligence translation software can grasp the general grammar concepts of a language based on the files it has learned before.

Speech and language technology researchers at Microsoft Research said that artificial intelligence translation technology does not use translation rules between languages, but rather treats translation as a problem and learns to translate between different languages ​​manually.

But Microsoft also acknowledges that translating news documents and translating human instant conversations are different. The latter uses more language idioms, and different accents and different dialects make it more difficult.

Last year, Google launched the wireless translation headset "Pixel Buds", which was dubbed as a "translation artifact" that can instantly translate 40 languages, but its functions and performance are still flawed, and non-mainstream language problems are also relatively large.

The dawn of artificial intelligence translation

Unable to learn

This is because there is not a large amount of translated text in non-mainstream languages, so artificial intelligence translation technology is learning nothing, such as translating Sinhala to Pashto. Problems such as this face greater problems. Challenge.

Of course, the translation software can translate Sinhala into English first, and then into Pashto, but this process is undoubtedly more prone to translation errors that have been said before.

According to Guillaume Lample, a Facebook artificial intelligence researcher, if enough translated text can be collected, there is a way to solve the problem of translation in small languages.

He said that the same reason, if the translation of small languages ​​can be resolved, then maybe humans can talk to aliens one day.

"We may learn to communicate with aliens one day, but first we must speak a lot and use the language we usually use."