【语义计算:议一议“微软机器翻译提前七年达到专业翻译水平”】 屏蔽留存

【语义计算:议一议“微软机器翻译提前七年达到专业翻译水平”】

屏蔽已有 1692 次阅读 2018-3-18 19:40 |个人分类:立委科普|系统分类:教学心得| NLP, NMT

 

李:
最近微信群疯传一条新智元的人工智能新闻专访,【机器翻译提前7年达到人类专业翻译水平,微软再现里程碑突破】。不少老友也特地转发给我这个“老机译”。微软这几天的营销好生了得。到处都是这个第一家超越人类的MT新闻 而且提前了七年!

这个微软 MT 是在哪里? 比较过百度,谷歌,有道。有道似乎最好,所以现在就用有道。不妨也试试微软。

事到如今,这几家都可以 claim 新闻翻译超过业余翻译的水平,进入专业翻译的段位了。跟语音类似,这是整个行业的突破,神经翻译大幅度超越上一代统计翻译,尤其是顺畅度。眼见为实,这一点我们都是见证人。已经 n 多次测试过这些系统了。(【谷歌NMT,见证奇迹的时刻】【校长对话录:向有道机器翻译同仁致敬】). 如果是新闻文体,很少让人失望的。

说是第一个正式超越的系统云云,基本上是 marketing。

MT 的下一个突破点应该是:(i)对于缺乏直接对比语料的语言对的 MT(据说进展神速);(ii)对于缺乏翻译语料的领域化 MT,譬如翻译电商领域,目前可用度差得一塌糊涂(20%左右),可有需求,无数据; (iii) 在保持目前NMT 目标语顺畅度的优势情况下,杜绝乱译,确保忠实可信。

这次他们严格测试的是汉译英,拿出数据来证明达到或超越了一般人的翻译水平。然后说,英译汉是类似的方法和原理,所以结论应该相同。这个我信。

有意思的是,在规则MT时代,绝不敢说这个话。汉译英比英译汉困难多了,因为汉语的解析比英语解析难,基于结构解析和转换的翻译自然效果很不相同。但目前的NMT 不需要依赖解析,所以语言的方向性对质量的影响很小。以前最头疼的汉译外,反而容易成为亮点。

当年入行的时候,有两个兴趣点:一是做外汉MT(主要是英汉),一是探索中文解析。前者有底气,知道这条路是通的,就是力气活,假以时间和资源,质量会逐渐越来越好。后者其实没有底气,感觉慢慢长路,“红旗不知道要打多久”(【从产业角度说说NLP这个行当】),但是实在太有趣了。当年的梦想是,什么时候中文解析做好了,用它来做汉外MT,能与外汉MT一样,那该多解气啊。

中文解析经过n多年的思索和实践,终于有底气了(【美梦成真】)。可是原先要落地MT的设想,却突然失去了这个需求和动力。好在 NLU 不仅仅在 MT 落地,还有许多可以落地的地方。

真所谓人算不如天算,看潮起潮落。老友谈养生之道,各种禁忌,颇不以为然,老了就老了,要那么长寿干嘛?最近找到一条长寿的理由,就是,可以看看这个世界怎么加速度变化的。今天见到的发生的许多事情,在 30 年前都是不可想象的:NMT,voice, image, parsing,iPhone,GPS, Tesla, you name it.

王:
中文解析,当然不一定中文,其他语言解析也一样,前景十分广阔,市场巨大。因现阶段还不能达到如人般的理解程度,所以还属于只能踩着有限小石子过河(落地)阶段,以后才能上大路,开高速。我也期望能有一个万能智能秘书,能准确理解且快速帮我办事,达到期望的结果。但能力都还有限,自然还是解析很力不足。

李:
parsing 的应用潜力很大,道理上没人说不对,毕竟这是AI在认知道路上可以预见的一个关键的支持。但实际上有两个坎儿:一是不好做,二是不好用。不好做是,想用的人往往不大会做,不能要求每个人都成为parsing专家。不好用是,独立的 offshelf 的,迄今没有见到大规模使用的成功案例。相对成功使用的,大多是内部消化,自己做自己用。这就局限了它的应用范围和潜力发挥。内部使用的成功经验,最多算是一种可行性论证,证明结构解析对于应用是的确可以赋能的。但平台化领域化的道路还很长。

核武器之所以在吆喝,是因为它还没爆炸,也似乎短期内不会爆炸。真爆炸了,听响声就够了,不需要吆喝了。

嘿,找到 MS Translator URL 了:https://www.bing.com/translator

做个现场测试和比较,用今天城里的新闻:

白宫发言人桑德斯14日表示,电视评论员柯德洛(Larry Kudlow)将出任国家经济会议主席。

桑德斯透过声明指出,川普向柯德洛提出担任总统经济政策助理,以及国家经济会议主席一事,柯德洛也接受了;白宫之后会宣布,柯德洛上任的时间。

川普月初宣布将对进口钢铁和铝材分别课征关税,其国家经济会议主席柯恩(Gary Cohn)因不认同川普此举,在6日请辞。

柯恩请辞后,柯德洛是接任的人选之一;在过去一年,柯德洛是川普的非正式顾问,也是川普2016年总统大选竞选期间的顾问。

川普13日向媒体表示,他有很高机率将任命柯德洛递补首席经济顾问柯恩辞职遗缺。川普说,“柯德洛成为我的好友一段时间了,他在我竞选早期就力挺我,是我最初的支持者之一。他是名非常有才华的人。”

70岁的柯德洛在前总统雷根任内监督白宫管理及预算局5年。柯德洛2001年加入财经媒体CNBC,之后成为知名评论家与主持人。

微软的自动翻译是:

White House spokesman Sanders said 14th that TV commentator Codro Larry Kudlow will serve as president of the National Economic Council.

Sanders said in a statement that Trump to Codro as president of the economic policy assistant, as well as the President of the National Economic Council, Codro also accepted, the White House will announce later, Codro time.

Cohn, the president of the National Economic Council, resigned on 6th for not agreeing to Trump’s move, announced earlier this month that it would impose tariffs on imported steel and aluminium.

After Cohen resigned, Codro was one of the successors, and in the past year Codro was Trump’s unofficial advisor and advisor to Trump during the 2016 presidential election campaign. Trump 13th to the media that he had a high chance of appointing Codro chief economic advisor Cohen to resign casual vacancies. Trump said, “Codro has been my best friend for a while, and he’s been my first supporter of my early campaign.” He is a very talented person.

” The 70-year-old Codro oversaw the White House administration and Budget Office for 5 years in former President Ronald Reagan. Codro joined the financial media CNBC in 2001 and became a well-known critic and moderator.

有道的译文是:

Larry Kudlow, a television commentator, will become President of the national economic council, White House spokesman jay sanders said Sunday.

In a statement, Mr. Sanders said he had accepted Mr. Cordelo as an assistant to the President’s economic policy and the President of the national economic council. The White House would then announce that Mr. Cordelo took office.

The chairman of the national economic council, Gary Cohn, said he would resign on June 6 after he announced that he would impose tariffs on imported steel and aluminum.

When cohn resigned, he was one of his successors. Over the past year, he has been an informal adviser to trump and an adviser to trump during the 2016 presidential campaign.

Trump told the media on Wednesday that he had a high chance of appointing the resignation of coderot’s chief economic adviser, korn/ferry. “He was a good friend for a while,” trump said. “he supported me early in my campaign and was one of my first supporters. He is a very talented person.

The 70-year-old has overseen the White House administration and budget office for five years under former President Reagan. Cordelojoined CNBC in 2001 and became a well-known critic and host.

谷歌的译文是:

White House spokesman Saunders said on the 14th that television commentator Larry Kudlow will serve as chairman of the National Economic Council.

Saunders pointed out through the statement that Trump had proposed to Cadero as the president’s economic policy assistant and the chairman of the national economic conference. Caldero also accepted it; the White House will later announce the time when Calderon took office.

Trump announced at the beginning of the month that it would impose tariffs on imported steel and aluminum products. Gary Cohn, chairman of the National Economic Council, resigned on the 6th because he did not agree with Trump.

After Keen resigned, Kedlow was one of the candidates for the replacement. In the past year, Kedro was an informal advisor to Trump and a consultant during Trump’s 2016 presidential campaign.

Trump stated to the media on the 13th that he has a high chance of appointing Kodlow to replace his chief economic advisor. Trump said, “Kadelo has been a friend of mine for some time. He supported me early in my campaign and was one of my initial supporters. He is a very talented person.”

Kodlow, 70, oversees the White House’s Office of Management and Budget for 5 years under former President Reagan. He joined the financial media CNBC in 2001 and later became a famous critic and host.

粗粗读一遍,都不错。可能是偏爱,还是更喜欢有道。

再来一片,其实不是新闻:

列宁的导师普列汉诺夫的遗言:
一、随着生产力的发展,知识份子的队伍比无产阶级增加得更快,在生产力中的作用跃居首位,在电气时代马克思主义的无产阶级专政理论将会过时。
二、布尔什维克的无产阶级专政将迅速演变成一党专政,再变为领袖专政。而建立在欺骗和暴力基础上的社会,本身就包含着自我毁灭的炸药,一旦真相大白,便会立刻土崩瓦解。
三、“布”党将依次遇到四大危机:饥荒危机、意识形态危机、社会经济危机和崩溃危机,最后政权土崩瓦解,这一过程可能持续数十年,但这个结局谁也无法改变。
四、国家的伟大并不在于它的领土甚至它的历史,而是民主传统和公民的生活水平。只要公民还在受穷,只要没有民主,国家就难保不发生动荡,直至崩溃。

微软:
Plekhanov’s last words:

With the development of productive forces, the team of intellectuals has increased faster than the proletariat, the role of the productive forces leaped to the top, and the Marxist proletarian dictatorship theory in the era of electricity will be outdated. Second, the Bolshevik dictatorship of the proletariat will rapidly evolve into a one-party dictatorship, then become a leader dictatorship.

A society based on deception and violence, in itself, contains self-destructive explosives that, once the truth is revealed, Fall apart immediately.

Third, the “cloth” party will encounter four major crises: Famine crisis, ideological crisis, social economic crisis and collapse crisis, the final regime fell apart, this process may last for decades, but this end can notchange. The greatness of the state is not in its territory or even its history, but in its democratic traditions and the living standards of its citizens. As long as citizens are still poor, as long as there is no democracy, the state will not be in turmoil until it collapses.

谷歌:
The words of Plekhanov:
     First, with the development of productive forces, the ranks of intellectuals have grown faster than the proletariat, and they have taken the lead in productivity. In the age of electricity, the Marxist theory of the dictatorship of the proletariat will become obsolete.
     2. The dictatorship of the proletariat of the Bolsheviks will rapidly evolve into a one-party dictatorship and will become a dictatorship of leaders. The society based on deception and violence itself contains self-destructive explosives. Once the truth is revealed, it will soon fall apart.
     Third, the “cloth” party will encounter four major crises in sequence: the famine crisis, the ideological crisis, the socio-economic crisis, and the collapse crisis. The final regime collapsed. This process may last for decades, but no one can change this outcome.
     Fourth, the greatness of the country does not lie in its territory or even its history. It is the democratic tradition and the standard of living of its citizens. As long as citizens are still living in poverty, as long as there is no democracy, the country cannot guarantee that there will be no turmoil until collapse.

有道:

His last words:
One, with the development of productive forces, intellectual team increase faster than that of the proletariat, in the role as the first productivity, at the age of electricity of marxism’s theory of the dictatorship of the proletariat will be out of date.

The dictatorship of the proletariat of the bolsheviks will quickly turn into a one-party state and become a leadership dictatorship. Society, based on deception and violence, contains self-destructive explosives, and when the truth comes out, it will fall apart.

Three, “cloth” party will, in turn, have four big crisis: hunger crisis, the collapse of the ideology, social and economic crisis and crisis, the regime collapse, this process may last for decades, but the end no one can change.

The greatness of a nation lies not in its territory or even its history, but in its democratic traditions and the living standards of its citizens. As long as the citizens are still poor, as long as there is no democracy, there will be no unrest until the country collapses.

梁:
这岂不让专业翻译人士好紧张吗?

李:
没办法,准备改行做编辑吧。作为职业,译员的市场的确在急剧萎缩中,大势所趋。

昨天跟老友说 还是要抬头讲故事 不能只是低头做实事。还要看故事的风口和时机。微软这一宣传 家喻户晓 老妪能解 普罗惊叹。一叹人工智能已经步步紧逼 看得见摸得着了。二叹微软太牛 给人印象是把对手远远抛在后面 在这个一日千里的AI时代 居然提前七年实现赶超人类语言认知的里程碑。

梁:
对,讲个好故事,比什么都重要!

李:
其实这几家品质都差不多 还有搜狗 还有一些初创 自从大约两三年前深度神经以后 都陆续达到了新闻翻译超越业余人工翻译的水平。换句话说 整个行业提升了。任何一家都可以心不跳脸不红做此宣称。可是老百姓和投资人不知道。这就看谁会讲故事了。

马:
大公司自己宣传,一帮不懂的媒体也愿意跟着捧,甚至捧得更卖力气。现在机器翻译拼的就是语料和平台,以前搜狗没有机器翻译,和我们实验室的刘洋合作后,不到一年就出了一个很不错的系统。

李:
AI 越来越像当年美苏的军备竞赛了,size matters.

@马少平 搜狗要营销的话 可以与电视台合作 搞个新闻现场大奖赛 请翻译界名人做评委 找n个专业翻译 m 个业余翻译 现场出题 限时翻译(要限制到熟练专业来不及查工具书 全凭大脑 勉强可以应付 为最佳)

马:
@wei 比起其他公司来,搜狗不是太会营销。

李:
这种比赛 没有悬念 最终一定是机器赢。好好设计一下,双盲比赛,让专家评审,也不能说它不公平。万一机器没得冠军 而是亚军或季军 宣传效果更佳 为下一轮比赛的高潮做了铺垫。那位人类选手神译 要好好保护 大力宣传 大书特书他过目不忘 博闻强记 知识渊博的种种事迹。当年花生智力竞赛大胜人类 本质上就是玩的这个套路。一直玩到进入了计算机历史博物馆的里程碑专区去了。MT 现在要玩的话,类似的效果,更容易设计,要想拉巨头参与也容易:几家巨头的MT网站都是公开的,随叫随到。

Ben:
@wei  youTube上 《成都》 有高圆圆的音乐,立委应该会喜欢!

李:
歌是好歌,早听过n多遍了,温暖慰藉。赵雷嗓子很有味道,可这小子镜头太多;圆圆友情出境,镜头太少,前面的剪影还是替身。

成都是个养人的好地方,出国前呆过大半年,乐不思非蜀(见【立委外传】):

1990 : 尝尽成都美食。 茶馆 火锅 夫妻肺片。

赵雷草根天才啊,独领城市歌谣,能写出这样的绝妙好词:

【画】
为寂寞的夜空画上一个月亮
把我画在那月亮的下面歌唱
为冷清的房子画上一扇大窗
再画上一张床
画一个姑娘陪着我
再画个花边的被窝
画上灶炉与柴火
我们一起生来一起活
画一群鸟儿围着我
再画上绿岭和青坡
画上宁静与祥和
雨点儿在稻田上飘落
画上有你能用手触到的彩虹
画中有我决定不灭的星空
画上弯曲无尽平坦的小路
尽头的人家梦已入
画上母亲安详的姿势
还有橡皮能擦去的争执
画上四季都不愁的粮食
悠闲地人从没心事
我没有擦去争吵的橡皮
只有一只画着孤独的笔
那夜空的月也不再亮
只有个忧郁的孩子在唱
为寂寞的夜空画上一个月亮

我不知道如何翻译,劳有道机器翻译一下:

“Draw a moon for the lonely night sky.
Draw me under the moon and sing.
Draw a large window for the cold house.
Draw another bed.
Draw a girl with me.
Draw another lace bed.
Draw a stove and firewood.
We were born to live together.
Draw a flock of birds around me.
Let me draw green ridge and green slope.
Picture peace and serenity.
The rain fell on the rice fields.
There’s a rainbow you can touch with your hands.
There are stars in the picture that I have decided not to destroy.
There are endless smooth paths.
The end of the family dream has entered.
Picture mother’s peaceful pose.
There’s also an eraser argument.
Paint food that is not sad in four seasons.
A leisurely person never worries.
I didn’t wipe out the quarrel eraser.
There was only one painting of a lonely pen.
The night sky was no longer bright.
Only a sad child was singing.
Draw a moon for the lonely night sky.”

自然有错译的地方(如 there’s also an eraser argument. I didn’t wipe out the quarrel eraser),可是总体而言,专业出身的我也不敢说一定能译得更好,除非有旬月踟蹰。机器翻译超越业余翻译,已经是不争的事实。

 

 

【相关】

【谷歌NMT,见证奇迹的时刻】

【校长对话录:向有道机器翻译同仁致敬】

【语义计算:李白对话录系列】

【置顶:立委NLP博文一览】

《朝华午拾》总目录

http://blog.sciencenet.cn/blog-362400-1104572.html

上一篇:【立委兄:小城记忆】
下一篇:【一日一parsing:修饰语的组合爆炸挑战】

 

3  黄仁勇 柳文山 徐令予

发布者

立委

立委博士,问问副总裁,聚焦大模型及其应用。Netbase前首席科学家10年,期间指挥研发了18种语言的理解和应用系统,鲁棒、线速,scale up to 社会媒体大数据,语义落地到舆情挖掘产品,成为美国NLP工业落地的领跑者。Cymfony前研发副总八年,曾荣获第一届问答系统第一名(TREC-8 QA Track),并赢得17个小企业创新研究的信息抽取项目(PI for 17 SBIRs)。

发表回复

您的电子邮箱地址不会被公开。 必填项已用 * 标注

此站点使用Akismet来减少垃圾评论。了解我们如何处理您的评论数据