Baidu’s AI can translate from one language to another in real time

The Chinese telecommunications giant Baidu has developed a translation system that may soon allow you to forget about the inconvenience of communicating with a native speaker who does not know. Baidu Research introduced the STACL platform , which can help with translation anytime, anywhere.

At the same time, the AI-technology, on the basis of which the translator is built, is able to predict what the interlocutor is going to say in the next few seconds, the system, if I may say so, “thinks ahead”. The work of STACL (Simultaneous Translation with Anticipation and Controllable Latency) can be compared to the work of the most efficient human simultaneous translators.

When a person is engaged in translation, he needs to listen at the same time, translate in his mind and speak, and all this is in a different language. This work is so complex that synchronists act as a team, replacing each other every 10-30 minutes.

If you take computers, they can easily perform all the necessary tasks, meringue breaks for rest and lunch. Until now, many features remained inaccessible to computer systems. The best of the best translation systems did what is called sequential translation. They listened to all that was said, translated it into the required format and, over time, issued a translation into another language. It all works well and correctly, but slowly.

Interestingly, in some cases, simultaneous interpreters will have to wait for key elements to appear. In German, for example, these are verbs - they go at the end of a sentence, not like in English. So, in order to understand what has been said, the translator has to be “in standby mode” for some time. In practice, this manifests itself in the form of pauses.

If we talk about the Chinese development, it is able to predict which verb should appear, so there are no pauses. The translation is very accurate - there are not too many problems with it. At the moment, STACL is still teaching, “feeding” computer newspaper articles. The result of the training is a fairly accurate prediction of how the narration will go on without knowing it.

Among other things, the Chinese system "understands" the specific features of the languages ​​with which it works, and uses it for translation. For example, simultaneous interpreters are sometimes mistaken when they are translated from Chinese into European languages ​​and vice versa - this is due to the specifics of the language itself. The machine is doing fine.

Of course, STACL also makes mistakes, and the robot has not yet been taught to apologize for them. A simultaneous interpreter can apologize for an unsuccessful attempt to translate, but the computer interpreter can simply be silent. At the moment, the system does not know how to notice and correct its own mistakes.

Nevertheless, the developed system can be customized, for example, to set speed and accuracy parameters - the faster the AI ​​translator works, the more it will make mistakes. If you need speed, you can put up with it (there are not so many errors, as far as you can understand).

In addition, the software platform can be further trained, “trained” on a specific topic. For example, STACL can be taught medicine, so that an electronic translator will practically make no mistakes while doing his work at a medical conference. “As in the case of any human translator, our system must be prepared to work before the event itself, where a translation is needed,” says one of the project representatives.

Introduce electronic translator for the first time at the conference Baidu World on November 1. At the event, the Baidu system will translate speeches by some speakers. The developers claim that STACL is not intended to replace human translators, especially in situations where accurate translation is crucial. In the end, no one wants to get to the center of an international conflict because the computer system was wrong and did not notice the error.

The purpose of the development is to provide users with a personal interpreter who is always with you. Previously, the company introduced a prototype of a user device that translates text on proposals with a new platform, presumably, the development of the gadget will accelerate.

Now STACL can translate text-to-text and speech-to-text translation. Developers want to add also the direction of speech-to-speech, so that everything is completely adult. Of course, in this case, everything is a bit more complicated, but the task is interesting, presumably, Baidu engineers can solve the problem.