Neural network machine translation (NMT) It has become one of the hottest topics in the localization industry , Compared with previous statistical based machine translation (SMT) comparison , It can improve the quality of translation 30％
, At the same time, the complexity of long distance language pairs is solved , Such as Chinese to English , Japanese to English, etc .
SDL According to a recent survey ,61％ Twenty five percent of respondents think machine translation is very important , However, it is not easy to find a machine translation solution suitable for the enterprise's own business . Today, Xiaobian was invited SDL AI
And vice president of machine learning solutions Mihai Vlad sir , Let's talk about the recent breakthroughs in machine translation SDL NMT The uniqueness of .
Q : comparison SMT,NMT The translation is more natural , Can you explain the difference between the two ?
Mihai：SMT and NMT
There is a big difference between the algorithm and the architecture of . Take autonomous driving as an example , The statistical method uses the driving data of specific road for training , Cars can drive well on specific roads . And the method based on neural network ,
Not bound to specific roads , Training using driving data from different roads , The car drives well on any road .
MT In fact, the development of 20 century
70 Era rule based machine translation , You can code a set of rules , But it will soon be realized that there are too many exceptions to different language pairs , Models are becoming more and more complex .1993
year , Machine learning is introduced into machine translation , The algorithm can be improved by learning bilingual corpus , Not through pre-set rules .
Q : The heat of artificial intelligence continues to rise , Every company claims to have AI technology . Machine translation is SDL One of the applications developed in artificial intelligence for many years , So how does the experience accumulated in machine translation help
SDL Develop other and AI What about its related applications ?
Mihai： Artificial intelligence aims to make computers copy human behavior . see , hear , get some action , Planning is typical human behavior , The most complex task is communication , The most difficult part is the ability to translate . about 40％
Of the world's population speaks only one language ,43％ You can speak two languages ,13％ You can speak three languages ,3％ You can speak four languages , only 1％ Can speak more than four languages .
We can move , run , Finally, we coordinate our actions to drive the car , But we can't master all the languages on earth . It is a very high requirement for the machine to have this capability . This is what artificial intelligence researchers will solve MT
The problem is seen as “AI complete” Why .
Q： Many consumer oriented online machine translation uses open source technology , Why should enterprises consider using enterprise level machine translation solutions ?
Mihai： And SMT comparison ,NMT
The code is more compact and complex , Some developers use open source projects , With hundreds of lines of code, you can create a translation tool that can learn from data . however , Enterprise solutions require hundreds of lines of code .
Scalable , Can be integrated and customized to improve quality , Only enterprise users NMT A few examples that need to be implemented in the code . Others such as , currency NMT
It is difficult to deal with document format well , Maintaining document format is a key requirement for any organization that wants to maintain document integrity ; General open source NMT Technology sometimes translates words repeatedly .
however , currency NMT One of the key issues is the cost ratio between training and translation SMT Several orders of magnitude higher . And companies that want to deploy this system may pay a lot of hardware .
In short , Using open source NMT system , You want a fluent translation system , The cost is very high , And there will be a lot of mistakes in the process of translation .
Technology