Machines, Lost in translation:the Dream of Universal Understanding (translated)

Original address: (2015-12-24)

http://www.npr.org/sections/alltechconsidered/2015/12/24/460743241/ machines-lost-in-translation-the-dream-of-universal-understanding?utm_content=buffera952d&utm_medium= Social&utm_source=twitter.com&utm_campaign=buffer

As early as 1954, scientists published for the first time a machine that could be translated between human languages. Later called the GEORGETOWN-IBM Experiment, the electronic brain can translate sentences from Russian into English.

Scientists believe that the world's translators, once developed, not only bring security to the United States on the Soviet border, but also raise world peace by removing language barriers.

They also believe that the evolution will soon be coming: Georgetown linguist, Leon Dostert, who started working with IBM's creator Thomas Watson, who believes that in the next five years or less, people will use electronic translation tools to build bridges between several languages.

This process indicates that it is quite slow. (very slowly, in fact, about 10 years later, the study's creator launched a survey of its lack of progress.) And more than 60 years later, a real real-time global translator-La C-3PO in Star Wars or the baby fish from The Hitchhiker's Guide to the Milky Way-is still a science fiction thing.

How far away is this realization? The opinions of the experts differ. As with many other areas of machine learning, it depends on how quickly a computer can be trained to simulate human thought.

Vikram Dendi said we were very close.

Dendi, chief technology and strategy consultant at Microsoft Research, said to all the technical people, "It's really cool to look back and say that we really turned science fiction into reality."

Microsoft's translation work has already produced apps that, in addition to the familiar text that translates into text, can turn sound into sound and sound into text. The biggest launch this year is Skype Translator, which records what you say in a video chat and can now turn them into voice or text translation in several languages.

Microsoft , of course, is not the only one. A company called Voxox to do internet phone chat, there is a text-to-text translation service, applied in its messaging app. Google , in addition to its very common text translation, has introduced a feature in its translation app, you can use your mobile phone camera to scan a foreign text picture and display translation.

promote the Machine brain

For decades, jumping language and technical barriers, scientists use a technique that is known as a neural network approach, in which machines are trained to simulate the way humans think-in essence, to create a human version of our Brain neural network.

neurons are nerve cells that can be stimulated by the human environment, including language. The longer a person is in an environment, the more complex the person's neural network becomes.

with the help of the neural network method, the machine converts each word into its simplest expression-a vector, equivalent to a neuron in a biological network, which includes information about each word, as well as the entire sentence or text. In the context of machine learning, neural network science has been developed for many years, and the more the neural network tries to translate, the more accurate results can be produced with limited manual help.

Professor of Computer science at the University of Montreal, which studies neural networks, says that although machines can now learn in the same way as humans, they still face some limitations. One of the limitations is the absolute amount of demand data--what children learn needs is far less than machine learning needs.

The language of the straggling

One of the challenges of implementing the global translator process is a very human one.

Some languages, although used by millions of of people, do not receive as much attention as computer science.

Hausa, for example, is a language used by 50 million of people in the West and Central Africa, but is considered a low-resource language because there is not enough translation documentation for computer scientists to use to harness machine learning.

Some scientists worry that these languages will be slowly disappearing, and that in this sense, machine learning will not work.

Jim Glass, a senior research scientist at MIT, said, "What is covered, and speaks, is now the biggest language." Until we solve the problem. C-3PO will not be a reality. "He is now studying whether machines can learn languages by interacting with real people."

In the Star Wars universe, the diligent C-3PO robot detects 6 million forms of communication, Glass says, "We can't even do 7,000."

Until then, he did not think scientists could really say that they were very close to having a global translator. --anne Li

"Machine translation requires a lot of computing and data; it doesn't make any sense," Bengio says. But the neural network approach is hopeful. It has the potential to achieve human level performance. It focuses on the meaning of words or conversations.

This method breaks (builds off) The method of the past machine translation.

In the early days, scientists taught computer translation by manually entering the rules for each language pair they wanted to translate. For example, in Russian, a noun followed by an adjective, the computer must know to flip it, in order to put adjectives in English in front of the simulation noun.

A press release detailing the 1954 GEORGETOWN-IBM trial says that translation between two languages requires more computer instructions than a simulated missile type.

In the face of the many rules and exceptions in each language pair, the manual input method quickly becomes tedious.

In the 1980, scientists began to move toward statistical-based models. The machine is fed back to many human-translated materials (for example, from the United Nations) and has their own language patterns and rules of identity.

Kevin Knight, a professor of natural language studies at the University of Southern California, says words that appear many times in a sentence are a common focus. "For example, by learning a large collection of English-Spanish documents, each time the computer sees Bancoon the Spanish side, you don't see the bank as bench in English. “

The computer will eventually infer that every time it finds a Banco de in Spanish, it can remove bench from its English options, since the Bank of general represents the name of a financial institution.

Testing Neural Networks

Neural networks, which became a popular tool for machine translation research in the 21st century, improved the quality of translation. The machine collects more information about each word and makes better odds analysis to avoid sounding unnatural translations.

What's the effect of this method? I decided to use Microsoft's Skype translator to try it out, this kernel is a neural network.

I am connected to Microsoft's Olivier Fontana via Skype video chat. Fontana greeted me in French, and a few seconds later, a male robot began translating his voice into English. To strengthen, I brought NPR's resident (French Pro) Caroline Kelly together. She says Skype seems to be more fluent in English-French than French-speaking English.

In the end, the results were surprisingly accurate, and were characterized by the discussion of typical kinship discussions, such as the summer vacation travel plan.

For any video conferencing, this translation chat relies on a strong network connection, which helps it extract laughter and weeding in the repeated ums and AHS.

Where translation becomes chaotic, we were discussing or trying to discuss the science behind Skype translator. The machine refuses to differentiate French words that correspond to "hip-hop" and "IPhone".

Handling spoken language in a sound-to-sound translation adds another layer of complexity to machine translation, because in addition to producing the correct results, the computer also needs to detect laughter, stuttering, repetition, and accents.

But, as scientists say, the more you use machine translation, the better they will become. "Neural networks are the creators of momentum," says Microsoft's Dendi. ”

"Without it, Skype translator is still a science fiction dream," Dendi said.

In other words, it is not possible to assert where machine translation will take place when the brain is like a human brain.

Machines, Lost in translation:the Dream of Universal Understanding (translated)