According to foreign media reports, Google released a web version and a mobile version of Google Translate yesterday. In the process of translation from Chinese to English, a new neural machine translation will be used, and the App will perform 18 million such translations a day. In addition, Google has published an academic paper on the operating principle of this translation system.
Earlier, Google had stated that they used neural network technology in Google Translate, but only limited this function to real-time visual translation. Some time ago, Google senior employee Jeff Dean told VentureBeat that Google had already tried to incorporate more and more deep learning functions into Google Translate. In addition, a Google spokesperson told VentureBeat in an email that the latest neuro-machine translation is the result of their efforts to develop deep learning capabilities.
In fact, Google has been working hard to integrate deep neural networks into its more and more applications, including Google Allo and Inbox by Gmail. This feature helps Google to process their data faster and more efficiently.
Google's Neural Machine Translation (GNMT) is highly dependent on eight-story short-term memory recursive neural networks (LSTM-RNNs). "Gradient flow can be enhanced through residual contact between layers," Google scientists wrote in the paper. With the help of the image processor, once the neural network becomes mature enough, Google can rely on its unreleased tensor processing unit for data processing.
Although neural machine translation is not always the best choice, Google's various attempts have shown that in some cases it is still exceptional.
“People’s evaluation of this translation system shows that compared to the previous phrase-based translation systems, the error rate of the neural learning translation system has been reduced by 60% when translating multiple languages, including English and French translations, and English and Western translations. The results of additional experiments show that the quality of translation systems will be closer to the average level of translators."
In a blog post published yesterday, Quoc Le and Mike Schuster, R&D scientists at the Google Brain Team, mentioned that with the help of bilingual scorers, the error rate of Google's neural machine translation was actually translated when translating multilingual sentences on Wikipedia. Has been reduced by 55% to 85%.
Despite this, this system is still not perfect. "Neuro-machine translations still make mistakes that translators will never make, such as missing words, correcting common names or rare proper nouns, lacking overall control over the context of the article, and so on. Therefore, we still have great room for improvement. But it is undeniable that neural machine translation has a milestone significance."
Interested readers can stamp out Google's original paper and original blog.
Via venturebeat
Cassava Dregs Separator,Cassava Processing Machine,Cassava Processing Equipment,Cassava Milling Machine
Hunan Furui Mechanical and Electrical Equipment Manufacturing Co., Ltd. , https://www.thresher.nl