Report of cloud of science and technology is achieved formerly.
"Congress of 2018 worlds artificial intelligence just rang down the curtain, division big question flew to be immersed in however " AI is passed together make a holiday " disturbance. On September 20, one passes dragoman to be in together know on dispatch says, the interpreter that dispatch flies to is passed to be the same as artificially actually, be not machine intelligence interpreter, and translation is undertaken by the machine recite, this matter caused the wide attention of media. This matter caused the wide attention of media..
Flying AI of division big question passed the be related that make a holiday to cause a mighty uproar together, while industry generates doubt with the ability that convey to AI, also let division big question fly into for public opinion center. Be aimed at this matter, flying Dongbi Jiang Tao responds to division big question external say, "Dispatch flies to never had said the idea that AI conveys together, emphasize the mode that is man-machine coupling from beginning to end. Emphasize the mode that is man-machine coupling from beginning to end..
Division big question flies so the man-machine coupling that the explanation travels a field together: "On interpreter caption Chinese and English when showing at the same time, it is embedded machine real time is full automatic pass together, when individual language shows, be artificial pass together. When artificial interpreter, dispatch flies it is OK to hear the help passes promotion efficiency and level together.
According to the explanation that division big question flies to, the machine that division big question flies to just rose to identify the speech that sends personnel together to translate, reoccupy AI sows interpreter machine the newspaper, concern with speech recognition, pass together with AI however have nothing to do.
In fact, division big question flies is not the enterprise of the first holiday that be built by doubt AI. Two years ago, a when search dog presses down show of world Internet congress in black " real time machine translation " the spot demonstrates, caused on the network " bogus science and technology " doubt.
Congress speech spot not only real time speech recognition shows Chinese content, still machine synchronism translates Cheng Yingwen, generate bilingual scroll caption. Respect of the dog that occupy search says, accuracy rate of its machine translation is amounted to 90% , prospective hopeful replaces the mankind be passed together.
However machine translation of search dog spot often appears " Chinese style English " interpreter, appear when Chinese identifying even when wrongly written or mispronounced characters, english interpreter is without a mistake however, the occurrence of these BUG can'ts help letting hostage doubt, do not become machine translation hard already hep manuscript, and be not a basis will real time Chinese undertake an interpreter?
No matter be Gu Ge, Microsoft, or Baidu, path, golden hill waits, domestic and international at present the accuracy rate to machine translation and the data with accurate neither one, basic with " translate wrong reduced rate " will measure its to progress, and search dog is cast external directly however gave the accuracy rate of 90% , make indeed the in an uproar inside course of study.
The black dragon farce of machine translation did not stop. This year in April, forum of Asia of rich a huge legendary turtle had 4 cent forum to apply Tecent AI to be passed together, result Tecent AI is passed together showed word chaos code, successive appear same the card machine state of a word. More serious is, tecent AI is passed together will " one belt all the way " changing the view that pattern translates into several kinds to differ, because this can urgent muster passes dragoman together,will save.
AI is machine translation actual level how?
No matter be division big question,the AI that fly sends the episode that make a holiday together, the onetime capital of machine translation thing that still is search dog, Tecent makes a person bemused, is the AI that is troubled by fervently all the time passed together in what administrative levels to stay? Does the distance send personnel together " by come off sentry duty " how long to return?
From the point of the show that at present AI conducts on congress together, sending personnel together still is the main force that sends a domain together at present, save to give accident Mr AI again and again even. In simultaneous interpretation domain, AI needs to face up to these things.
Does the accuracy rate of machine translation have AI much taller?
Evaluate the stand or fall of quality of AI machine translation content, there are a few methods on international, besides artificial evaluation, also have a variety of automatic evaluation methods, among them, the Watson research center of IBM is approbated relatively extensively by industry in the BLEU means that offerred 2002.
Referenced " machine translation and brief introduction of its evaluation technology " view, BLEU method compares Hou Xuanyi article namely (Candidate) with referenced translation (Reference) in same extract amount. Its thought is the N that in using interpreter result, appears continuously yuan group (N word / word or punctuation) with the N that appears in referenced translation yuan group undertake comparative, the N that computation matchs completely yuan the N in a number of the group and interpreter result yuan of the group total a scale of several.
Next, the author uses corpora of a test, let 2 interpreters expert and 3 machines translate, namely it allows textual have many referenced translation. But remember please, to the absolute value of accuracy rate neither one of machine translation, however with " translate wrong reduced rate " the maturity that will judge oneself technology.
Do not pass to AI word, if want to raise its accuracy rate, must assure it above all you can " understand " and " listen thoroughly " .
Had been machine translation technology examined?
Before search dog machine translation is mocked, another reason is without test and verify, this among them the biggest BUG is the evaluation that the technology has not obtained sufficient capacity degree, namely technical maturity without market test and verify. Have medium for instance with respect to speak bluntly " from only then to eventually, the user that search dog real opportunity turns over has Wang Xiaochuan only one person. The user that search dog real opportunity turns over has Wang Xiaochuan only one person..
So " light says not experienced false skilled person " , the technology cannot be regarded as on the wall by the hammer Bodhisattva is done obeisance to, AI should be replaced truly artificial, do the human Gao Fu such as simultaneous interpretation, writing's miscellaneous work, have to apply in actual combat to come first.
How is the machine short of the affection element that miss to fill?
The language is door art, it is vivid, it is to have rise and fall sentient, put below a large language area only ability is likely by completely accurate understanding. Even if is same a word, what convey below the condition is being put in different language also is different emotional colour, corresponding interpreter is different also.
Perhaps, AI machine can achieve 95% accurate interpreters, but that leaves the feeling of 5% , how should convey again come out? This affection of 5% creates certain understanding obstacle possibly is misunderstanding even.
Want to let these machine expression give these affection elements, most fundamental premise is AI needs to improve words condition savvy. At present intelligence translates complex to Chinese meaning sentence the tense in type discerns capacity is insufficient still, the meaning talent that often needs sufficient understanding is full word is accurate hold tense.
But intelligent interpreter is not be based on understanding, be based on however statistical with the corpus, can accomplish only translate this word meaning roughly accurate, identifying emotional colour field so, still need to strengthen.
Does spoken language change can special processing?
Say in front, want to raise the accuracy rate that AI translates, need assures can " understand " and " listen thoroughly " , this asks can accurate judgement gives AI to be translated not only object (like lecturer) speech, halt, undertake inside extremely short time " speech make pauses in reading unpunctuated ancient writings " , still need to be in when involving a few ambiguous sound can basis " context " undertake be adjustmented in time, more intractable is to should make A can agile handle a few special cases, for instance spoken language.
Want to handle these complex cases, AI needs a large number of compensatory spoken language to change corpora, and need change metaphrase kind, make machine translation true intelligence rises.
In fact, simultaneous interpretation needs to extremely tall qualification just can be held the position of. In simultaneous interpretation, dragoman needs to abide by arrange a drive, adjust at any time, measurable lead, information recombines, the basic principle such as equivalence of reasonable and contracted, information, in addition, because lecturer speech rate is general faster, each district accent and dialect may adulterate in the speech, all intellectual reserve that pass dragoman to still need to bring his into play together and experience hammer and tongs.
Look the high-key catchword that replaces the mankind to be passed together with the machine is in short-term inside still be theoretic only. What system of nerve machine translation can do at present is to optimize interpreter result only, not be elixir, regard black science and technology as particularize its, unplug overly high people is real to machine translation horizontal expectation, cannot take actually.
It is division big question flies not just, dog of Tecent, search, all artificial intelligence application and company should face up to the difference with the mankind and real machine, this is a technology not just who actors or actress the argue of whose bad, it is the stand facing each other of a requirement that how understands person and person more. If company of science and technology gives people too much highbrow illusion, so virtually also will overdraw the public the interest to AI.
[report of cloud of science and technology is achieved formerly]
Date of Zhang of small letter public: Cloud of science and technology reports
