Google launched the translation of neural networks in the translation of the level of human life – S

Google launched the translation of neural networks in the translation of the level of human life? – former news translation model of Sohu is a phrase type, a word translation system will be split into word or phrase translation, often get some results of mere trash. While the new translation system called Google Machine Translation (GNMT), the nerve of this new machine design technology with less choice, it will be a sentence as a whole is decoded, though still a sentence is divided into several independent units, but each unit on the sentence before and after consideration. It’s like a junior high school entrance to high school, to complete the assignment, to try to find words and phrases in a huge dictionary. Nerve Machine Translation and previous translation mode is also a big dictionary, but the old mode of translation just one by one word search, while Machine Translation has trained two neural networks, can produce another language full text. This is very important for the Chinese translation, because in the Chinese language because of the difference in a word often have different meanings. Why can Google translate a qualitative leap in accuracy? It is reported that Google has overcome before Machine Translation translation nerve nerve in accuracy and speed of the defect, with 8 layer and 8 layer encoding and decoding the length of memory (LSTM) network to enhance attention and record the moment of feeling left. The mechanism of attention is attached to the lower layer and the upper layer of encoding and decoding, in order to improve the degree of parallelism in order to reduce the consumption, improve the final translation algorithm speed low precision in calculation of reasoning. The first step, the network will Chinese word encoding a series of vectors, each vector represents the sentence meaning of all the words. Once the whole sentence is read by the neural network, the decoding begins, and the corresponding English sentence is generated. Decoding is the process of the weighted distribution of the encoded Chinese vector and the related English words. A curve representing the encoding and decoding of the encoded word. In order to enhance the treatment of rare words, Google will words into sub word units limited, which is convenient to input and output. The beam search technology to search length normalization and a cover with all the words, which makes the output pattern is covered by a large amount of language translation. All of this can be classified as artificial intelligence, artificial intelligence algorithm does not rely on human logic, compared to the previously used manual algorithm, artificial intelligence algorithm can find a better way to complete the task. The artificial intelligence network itself will learn how to translate, it is only focused on how the results, and not subject to human interference. Mike, a Google engineer who developed the project, said, "you don’t have to choose, the system will be fully translated."." Only one of the many languages supported by Google translation. Google said that the future will rely more on the translation of artificial intelligence. In the next few months, Google will also be designed for more language Google nerve Machine Translation. Of course, this kind of nerve Machine Translation can not be regarded as perfect, of course, based on artificial neural intelligence performance is good, but there are still a lot of subtle machine algorithm can not really understand. Compared to human translation, Google neural network.相关的主题文章: