Progress is a terrible "mechanical translation", and ask the top engineer about the reason (Diamond Online) --Yahoo! News

Photo: PixTa

　The progress of machine translation such as Google translation and DEEPL is tremendous. What kind of technological innovation did you have? How does communication change as accuracy goes up? Is there a day when Japanese people don't need to study English? NTT Communications Basic Research Institute for long -term mechanical translation, Susumu Akiyama, a special researcher of the Linguist Institute for Group, Susumu Nagata, and Susumu Akiyama, the author of this series, "Organizational Disease". We talk about "Promotion of Machine Translation and Future". In the first part, we explained the development of machinery translation and the revolutionary technological innovation of neural networks in detail and in an easy -to -understand manner. (Interview/composition/Writer Okuda Miki Okuda) [See the image of this article] ● The same technology as the hedge fund started in Akiyama 4-5 years ago, and the level suddenly rose. I thought. And two or three years ago, the automatic translation service "Deepl" came out, and I thought it would not have to be studying English if this was already there. Please tell us the background of the rapid level of the level in recent years and the change in machine translation until recent years. Looking back on Nagata history, the computer principle of computer, which would have been used for cryptographic decryption during the war, was created. It comes out immediately after the war. During the Cold War, when the conflict between the United States, the US Department of Defense has created a project to translate Russian into English to thoroughly study Russia. With a large amount of investment, automatic translation has begun in the "rule -based" method. When she says, "She ate a cake (she is She ate The Cake)", Japanese is a subject, object, verb, English is S -subject, V verb, and O object. You can see that there is a rule that the verb and the order of the object should be turned over. And she makes a biling dictionary for individual words such as = she, ate = ATE. Analyze the input syntax structure (such as the order of elements such as subjects, objects, verbs, etc.), apply the conversion rules of the word order (turn the verb and object overturn), replace the dictionary, and translate it. Implemented in. In this way, you can do automatic translation in principle. However, human language expressions are diverse, so no matter how much the rules are added, they cannot catch up. As the rules increase, the limits of the rules that can be managed come. The limit of the number of human management rules is about 1000. If you add a new rule, the accuracy will not increase and contradiction will occur. However, humans have worked hard for 50 years. Did Akiyama Nagata also worked hard (laughs)? Nagata: In the next era of the rule base, I participated since the 1990s "statistical translation". No matter how much you do on the rule base, the results of the Watson Research Institute in IBM in New York have started "Statistical Machine Translation (Statistical Translation)". It is a mechanism to collect a lot of pair data in the translated language, statistical analysis, and learn what is equivalent to conversion rules and bilingual dictionaries. This went well. However, in 1991 and 1992, the United States was hit by a great recession, and IBM was forced to significantly reduce the number of employees. Researchers flowed to Wall Street, and a company called Renaissance Technologies, which became famous as hedge funds, called a high -frequency trading (high freak ency. trading), and the computer instantly determines the stock price and volume, etc. We have created a program for trading trading by automatic order. The core is an algorithm that predicts the stock price after a few seconds after the comma is the same as the statistical translation of what is changing in chronological order. Researchers got huge wealth on the walecular streets. However, the statistical translation itself has soon disappeared.

Page 1/5