Microsoft technology academician Huang Xuedong: Machine Translation will reach human professional level

category:Internet
 Microsoft technology academician Huang Xuedong: Machine Translation will reach human professional level


In less than two years, Huang Xuedongs team made breakthroughs one after another. His achievements were also recognized by Microsoft. In 2017, he was selected as the Microsoft technology academician (TechnicalFellow), representing the highest honors of Microsoft technicians, which also included the Turing prize winner ButlerWLampson, CharlesP.Thacker and so on. Huang Xue Dong Recently, at the Microsoft Asia Research Institute in Beijing, one of the architects of the IEEE and the ACM (American Computer Society) academician, the Microsoft cognitive toolkit, CNTK, accepted an interview with the www.thepaper.cn. From the university time to joining Microsoft, taking 25 years of technical journey and to lead the team to a lot of success, Huang Xuedong talked about why he had been able to stick to it for a long time: it was to make language communication better, and the language barriers between people and people could be eliminated. At present, Huang Xuedongs team has reached the level of professional stenographer in the speech recognition rate of speech recognition, and is on the same side of Machine Translation professional translators. In the interview, in addition to the story behind the latest Machine Translation breakthrough, Huang Xuedong also talked about whether the development of artificial intelligence will eventually replace human beings, as well as the debate on artificial intelligence between China and the United States. Artificial teacher: Microsoft Machine Translation surpassed human translators 7 years ahead of schedule. In 2017, University of Oxford had done a survey of MachineLearning researchers to predict the future of artificial intelligence. According to the prediction of these researchers, artificial intelligence in the next 10 years will surpass mankind in many ways. Among them, they predict that Machine Translation will need more than 8 years of human amateur translators. According to this forecast, the Machine Translation system developed by Microsoft in March this year actually advanced the time of 7 years ahead of the human amateur translators in Machine Translations translation of the English language and reached the level of human translators. This is a historic breakthrough, because language has always been the most important thing for us to communicate. In particular, China is going to go to the world. It is now possible to reach a professional level. This is a historic breakthrough. This is positive and inspiring for China to move towards the world and promote artificial intelligence. Huang Xuedong told the www.thepaper.cn. According to Huang Xuedong, the project was made up of three research groups from the Microsoft Asia Institute and the Microsoft Lede Meng Institute, which carried out joint innovations across the Sino US time zone and across the research field. Since 1954, Machine Translation has been a field for human beings to tackle. But in the past few decades, researchers have been unable to find effective breakthroughs because of technical methods, computational power and training data. It was not until twenty-first Century that researchers introduced the neural network of deep learning to build a breakthrough in Machine Translation (NeuralMachineTranslation). ? Neural Machine Translation, in brief, is to encode the language of the source language, which is translated into the form of understanding by the computer, and the result of the encoding will form many hidden variables, each of which represents the semantic information of the clause from the first to the current vocabulary. Then, through a decoding process, a word and a word are output. In Microsofts innovation, Microsofts research team drew on some of the ways of thinking that humans translate. The first is dual learning (DualLearning). It is a little like that when people are learning English, when they are translated into English, they usually need to be checked, and then the English will be translated back to Chinese to see if it is correct and achieve the consistency of the translation. The purpose of this is to achieve joint optimization in the process of Chinese English and Chinese translation. Microsoft has taught the method to the machine. The second is to deliberate the Internet (DeliberationNetworks). In Chinese poetry, the knocking of monk knocking on the moon door is a result of the authors repeated attempts to write poetry and express the language to the extreme. In Microsofts Machine Translation system, it has learned from this way. After Machine Translation completes the translation, in order to make it better, the researchers use the network to make the machine constantly revise, constantly modify, reenter the result of the translation, and then optimize it continuously. The third is consistency specification (AgreementRegularization). Consistency specification allows translation to be done from left to right, or from right to left, resulting in two processes generating consistent translation results. Huang Xuedong explained that this is like the process of human translation. We usually look left and right, and both sides are pleasing to the eye. In addition, the Machine Translation system has also used JointTraining, which can be considered a combination of learning from source language to target language translation (SourcetoTarget) and learning from target language to source language translation (TargettoSource). Both Chinese and English translation and English and Chinese translation use initial parallel data to train. During each iteration of the training, the Chinese English translation system translates Chinese sentences into English sentences, thus obtaining new pairs of sentences, which can also be added to the data set of the English and Chinese translation system. In the same way, the process can also be carried out in reverse. This two-way fusion not only greatly increases the training dataset of the two systems, but also greatly improves the accuracy. It is worth mentioning that the Microsoft research team has also found bilingual language consultants for the whole translation study and evaluated the results of Machine Translation. As people like learning English, they usually need a professional teacher to evaluate the results of our study. For Machine Translation, the results of a professional teacher are also very important. Machine Translation is actually equivalent to a question with more than one answer, the answer is good and bad with the teachers level, so its hard to judge by machine automatic scoring. We now use professional staff to evaluate the quality. Huang Xuedong said. Under the above technologies, Microsofts research team set a two - year study cycle and completed the research results of Machine Translations professional level one year ahead of time. The edification of three different universities and the persistence of the last 30 years Huang Xuedong has been in Microsoft for 25 years, and he has been studying in the University for more than 30 years in the field of artificial intelligence. From the first unknown to the last few years, the team led by himself has completed a breakthrough, and Huang Xuedong said it was a persistent achievement and a lucky one. Yuelu Academy has more than one thousand years of history. At the same address at Yuelu Academy, Hunan University has maintained the continuity of education. Every day I pass the Yuelu Academy from my dormitory, I want to see the couplets in front of Yuelu Academy. It lays the foundation for self confidence and emotional pursuit. Because a lot of things and your pursuit of career must have endurance, internal force, this is the best training that Hunan University gave me. Huang Xuedong said. But it is the change of learning environment that makes Huang Xuedong find the way he wants to study. 30 years ago, I studied American English in Tsinghua, studied in Britain, and went to Edinburgh, Scotland. Although the school was very beautiful, it was also the place for the cow. But I cant understand in class because my English is not good enough, plus Scotland accent. So I wanted to talk about how good the subtitles are. Huang Xuedong smiled and said, at that time, Britain is still very powerful, 1987, all TV programs are subtitled BBC, I was transferred to the Scotland local channel to watch the English subtitles in Scotland, I now speak English with Scotland accent. Deeply aware of the pain in his learning process, Huang Xuedong hopes to help others solve this problem. Much of the work and research he has done in Microsoft is also related to these experiences. For example, Microsofts voice translation, they have Skypetranslator, Microsofttranslator two landing products. At present, Microsofttranslator can be used in both iPhone and Android devices, and is integrated with Microsoft Office. This app can communicate with face to face with language barriers while supporting 100 people to communicate in more than 60 different languages. ? After 30 years of persistence, Huang Xuedong said his ultimate dream was to make machines understand human language and allow machines to speak and communicate like humans. At present, his team has achieved success in speech recognition and Machine Translation, and the next goal is to make breakthroughs in speech synthesis. The competition pattern of China and the United States artificial intelligence From Chinas universities, Huang Xuedong has studied and worked in Britain and the United States. As a global technology giant, Microsoft has made an in-depth layout in the field of artificial intelligence. The Microsoft Asia Research Institute in Beijing is also known as the Whampoa military academy, which is called artificial intelligence talents training. Huang Xuedong has his own view on the increasingly obvious competition between us and China in the field of artificial intelligence. In his view, AI wants to be successful without three elements: computing power, data and algorithms. From these three dimensions, China has a large number of users, data and advantages in data, which is a bright spot in China. But there is only data, no computing power and algorithm is far from enough. Dont think about it. Its a good thing to have a good reputation. The achievement of any achievement is not only a personal achievement, but a step forward from the research and development of generations. This is a fact. The whole community is not a Microsoft company. It is a joint effort of university research institutes and companies. Huang Xuedong said.