Share on LinkedIn There is a well-known AI (Artificial Intelligence) phenomenon, called Eliza Effect, stating that people may over-interpret the machine results, reading between lines for meanings that do not originally exist. Here is the entry in Wikipedia: The ELIZA effect , in computer science , is the tendency to unconsciously assume computer behaviors are analogous to human behaviors. In its specific form, the ELIZA effect refers only to the susceptibility of people to read far more understanding than is warranted into strings of symbols — especially words — strung together by computers. ...... More generally, the ELIZA effect describes any situation where, based solely on a system's output, users perceive computer systems as having intrinsic qualities and abilities which the software controlling the (output) cannot possibly achieve or assume that reflect a greater causality than they actually do. ...... The discovery of the ELIZA effect was an important development in artificial intelligence , demonstrating the principle of using social engineering rather than explicit programming to pass a Turing test . ( https://en.wikipedia.org/wiki/ELIZA_effect ). In fact, for human intelligence, there also exists a mirror effect, what I name the Anti-Eliza effect, which relates to the tendency to unconsciously mythify human capabilities, by often over-interpreting output of human agents for meanings which do not originally exist. The Anti-Eliza effect disregards the fact that more than 90% of the human intelligent activities are actually mechanic or algorithmic bu nature, supported by access to the memory of a knowledge base. In fact, the frequently observed Eliza effect and the Anti-Eliza effect are two sides of the same coin, based most likely on similar cognitive grounds of human mind. The human intelligence in effect can hardly stand undergoing decomposition for a few rounds before it shows its true identity of possibly less than one percent of inspiration, with 99% mechanical processes. When they are blended together, they may manifest themselves inside a human body to be worshiped as a master or genius. There is no way for the Artificial Intelligence to target that one percent. It is neither possible nor necessary. Hereby let me present this new concept of the Anti-Eliza effect in AI, to be associated with the human habit and nature of self-mythification. Such human self-mythification is exemplified by reading the human intelligent output for meanings which simply do not exist and by over-exaggerating the significance of human spirituality. For example, for the same piece of work, if we are told the work is a result of a machine, we will instinctively belittle it, in order to maintain the human dignity or arrogance. If the work is believed to be a rare antique or artifact of a human artist, it will draw numerous interpretations with amazing appreciation. The Anti-Eliza effect shows itself widely in the domain of art and literature review. For the genre of abstract art, this effect is rationalized: it is actually expected for different people to read different meanings out of the same art, independent of what the original author intended for. That is considered to be part of the original value of this type of work. The ability of reading an artistic work for many meanings which were not intended for is often considered to be the necessary quality of a professional art reviewer. It not only requires courage but also is often futile to point out that the emperor has no clothes on and the work does not make sense, or has no meanings as interpreted by reviewers. The theory of aesthetics involving the abstract art has no needs to depend on reality checking at all. In my understanding, the Anti-Eliza effect is manifestation of mysticism, and mysticism is very close to some nature of human beings. This is a huge topic, that calls for further exploration in AI to see the entire picture and scope of this effect. The Anti-Eliza effect is believed to be an important basic concept in the field of AI, as significant as its mirror concept of the Eliza effect. This is by no means to deny the supremacy of human mind, nor to deny the humanity shined by that one percent of spirituality in our intelligent output. Only the most stupid people would be so self-denial, attacking human dignity. However, for either science or engineering, everything needs to be verified or proved. In the AI work, the first thing we do in practice is to peel off what can be modeled by a machine from what is truly unique to humans only. We will then make the machine mimic or model that 99% of materials or processes in human intelligent behaviors, while keeping a distance from, and maintaining a high regard for, the 1% true human wisdom. As is observed, the routine intelligent activities of mankind will be approximated and modeled more and more in AI, very much like a silkworm eating more and more parts of mulberry leaves. With each progressive territory expansion of AI, what was originally thought of as truly intelligent is quickly decomposed into an algorithm of solutions which no longer belong to the human unique wisdom. If the nature of mankind is simply a hybrid of 1% from the holy spirit and 99% from some fundamentally mechanical devices, then in the end, it is inevitable that machines will one day replace the 99% of human intelligent labor. From a different angle, any implementable AI-oriented proximation to the human intelligent activities is by nature not true intelligence. Obviously, as time goes by, more and more such proximation will be programmed into machines to replace the mediocre human performers. But there will always be something that can never be computerized, which is the true human intelligence (synonyms of this include wisdom, spirituality, inspiration, soul, etc). The difficulty now lies in that for majority of specific intelligent tasks, they are still mixed together with no clear separation between spirit and mechanical materials. We can hardly define or see clearly what that spirit (the core intelligence unique to mankind) is, unless AI accumulates its modeling successes over time to a point of diminishing return. Before that happens, we humans tend to continue our mythification of our own abilities, and classifying those abilities as uniquely human. In other words, the Anti-Eliza effect will run a long long time based on the human nature of mythification. Let us look at the history for a brief review of some fundamental abilities which have long been believed to be human intelligence. This review will help us see how this effect has evolved over time. In the pre-computing era,the arithmetic abilities were highly regarded. The few people with exceptional arithmetic performance were generally considered the most intelligent men. After calculators and computers were invented, this was the first myth to break down. No one in today's society will consider a calculator an intelligent being. Following the calculating power is the memorization capacity that has also been believed to be an incredible intelligence of the human brain for a long time. In Ancient times, people with extraordinary mental arithmetic ability and outstanding memory capacity were often worshiped as genius masters or living gods (of course, memorization involves not only the storage capacity, but also the accompanying retrieval abilities to access the storage). As a matter of fact, many intelligent machines (e.g. some expert systems) implemented in the AI history come down at the core to a customized search algorithm plus a memory of formalized domain knowledge. The modelled intelligent activities have thus been demystified from the presumed Anti-Eliza effect. For an illustration, I would like to present the case of natural language parsing to see how much human intelligence is really involved. The ability to parse a natural language in grammatical analysis is widely recognized in the NLP community and beyond as a key to the natural language understanding and the human intelligence. Fortunately, modeling this ability has been one of the major professional tasks in my entire career in the last two decades. So I believe that I have the expertise and knowledge to uncover the true picture involved in this area. As a seasoned professional, I can responsibly tell you, 99% of the human parsing capability can be modeled by a computer very well, almosy indistinguishable from a human grammarian. The human grammatical analysis of the underlying linguistic structures can be closely assimilated by a linguistic parsing algorithm based on the language knowledge of vocabulary, usage and rules. Both our English parser and Chinese parser, which I designed and led the team to have developed, are close to the level of being able to parse 99% of random text into reasonable linguistic structures. As a key human intelligence, this type of modeling performance was unimaginable, like a miracle, but there lie some easily measurable benchmarks in practice. Stepping back from parsing abilities to the metaphysical level, what I want to say here is that, much of what we originally thought of as deep intelligence cannot stand decomposition either. Every step of decomposition in the AI research progress has helped to reveal the true picture of human intelligent behaviour which is usually not what we used to believe. It has turned out to be a wonderful and eye-opening experience in the career of most AI and NLP researchers in the last few decades. Until things are decomposed in AI, there has been a natural Anti-Eliza effect that seems to control or dominate the perception of most types of human intelligent activities in the minds of not only the general population but us AI-insiders as well. Before we embark on exploring an intelligent task from the AI perspective, we often cannot help mythifying the seemingly marvelous intelligence due to the Anti-Eliza effect. But most of us end up with being able to formalize a solution that is algorithmic and implementable in computer with a memory of some form of knowledge representation. We will then realize there is really little magic in the process. Most probably, I believe, mysticism and self-mythification are just part of human nature, hence the widespread manifestation of the Anti-Eliza effect. translated by the original author Wei Li and from his original Chinese version here: 【新智元笔记:反伊莉莎效应,人工智能的新概念】 【相关】 【新智元笔记:反伊莉莎效应,人工智能的新概念】 MT 杀手皮尔斯(翻译节选) 【置顶:立委科学网博客NLP博文一览(定期更新版)】
我们 “语义计算” 群在讨论这个句子的句法结构: The asbestos fiber, crocidolite, is unusually resilient once it enters the lungs, with even brief exposures to it causing symptoms that show up decades later, researchers said. 我说,it looks fine in its entirety. once-clause has a main clause before it, so it is perfectly grammatical. The PP with even brief exposures to it is an adverbial of causing ...: usually PP modifies a preceding verb, but here it modifies the following ING-verb, which is ok. 然后想到不妨测试了一下我们的 parser,果然,把 PP 连错了,说是 PP 修饰 enters,而不是 causing。 除此而外,我的 parse 完全正确。这也许是一个可以原谅的错误。如果要改进,我可以让两种可能都保留。但是统计上看,也许不值得,因为一个 PP 面对前面的一个谓语动词和后面的一个非谓语动词,修饰前者的概率远远大于修饰后者。 张老师问: 是否此句在你的训练集里?如是统计方法。否则太不容易了 我说,我这是语言学程序猿做的规则系统,不是统计方法。句子不在我的 dev corpus 里面。parsing 是一个 tractable task,下点功夫总是可以做出来,其水平可以达到接近人工(语言学家),超越普通人(非语言学家)。说的是自己实践的观察和体会。靠谱的 parsing,有经验的语言学程序猿可以搞定,无需指靠机器学习。 为了说明这个观点,我测试了我的汉语 parser: 这个汉语句子的 parsing,只有一个错误,“语言学”与 “程序猿” 之间掉链子了(说明 parsing 还有改进余地,汉语parsing开发晚一些,难度也大一些,当前的状况,掉链子的事儿还偶有发生)。但整体来看基本也算靠谱了。所以,即便是比英语句法更难的汉语,也仍然属于 tractable 人工可以搞定的任务。 语言学家搞不定的是那些千头万绪的任务,譬如语音识别(speech recognition),譬如文章分类 (document classification),譬如聚类习得 (clus tering-based ontology acquisition) 。这些在很多个 features 中玩平衡的任务,人脑不够用,见木不见林。但是对于 deep parsing 和 信息抽取,解剖的是一颗颗树,条分缕析,这是语言学家的拿手好戏,都是 tractable 的任务,当然可以搞定。(甭管多大的数据,一句句分析抽取完了存入库里,到retrieve的时候还是需要“挖掘”一番,那时候为了不一叶障目,自然是需要用到统计的)。 在 条分缕析的 tractable 任务上(譬如,deep parsing),我的基本看法是:有NLP经验的语言学家立于不败之地。而机器学习,包括深度学习(deep learning,当前呼声最高的机器学习神器),也许在将来的某一天,可以逼近专家水平。值得期待。最多是逼近语言学家,但是要超越人工,我不大相信。再牛的机器学习算法也不可能在所有任务上胜过专家的手工编码,这个观点本来应该是显然的,但是学界的多数人却天然地认为深度学习总是可以超越人工系统。 parser 的直接目标不是语义求解, 而是提供一个靠谱的结构基础, 这样后续的(语用层面的)语义理解、信息抽取、舆情分析、机器翻译、自动文摘、智能秘书或其他的NLP应用, 就可以面对有限的 patterns, 而不是无限的线性序列。 从这个目标来看,我们的中文英文的 parsers 都已经达标了。 【相关】 【围脖:做 parsing 还是要靠语言学家,机器学习不给力】 手工规则系统的软肋在文章分类 《立委随笔:语言自动分析的两个路子》 再谈机器学习和手工系统:人和机器谁更聪明能干? 【 why hybrid? on machine learning vs. hand-coded rules in NLP 】 Comparison of Pros and Cons of Two NLP Approaches 【置顶:立委科学网博客NLP博文一览(定期更新版)】
相关论文和资源:(cvpr13 cvpr14 ijcv15) Scene Parsing with Object Instances and Occlusion Ordering J. Tighe, M. Niethammer, and S. Lazebnik, CVPR 2014 Project webpage , IJCV article Finding Things: Image Parsing with Regions and Per-Exemplar Detectors J. Tighe and S. Lazebnik, CVPR 2013 Project webpage , IJCV article 待续
上次提过,先搜后 parse ,是可行的。 早在十几年前,A skJeeves 被华尔街追捧。这里面也有很多 IT 掌故 我专门写过博文( 【问答系统的前生今世】 , 【 金点子起家的 AskJeeves 】 )。 当时NLP (Natural Language Processing) 红透半边天,下一代 Google 呼之欲出的架势,尽管A skJeeves 其实 NLP 含量很低。他们不过利用了一点 NLP 浅层对付问题的分析。这才有我们做真正基于NLP的问答系统的空间。 就在A skJeeves 上市的当天,我与另一位N LP 老革命 Dr. Guo ,一边注视着股市,一边在网上谈先 search 后 parse 的可行性。此后不久我的团队就证实了其可行,并做出了问答系统的 prototype ,可以通过无线连接,做掌式 demo 给投资人现场测试。当年还没有 smart phone 呢,这个 demo 有 wow 的效果,可以想见投资人的想象被激发,因此我们顺顺当当拿到了第一轮一千万的华尔街风投(这个故事写在 《朝华午拾:创业之路》 )。 问答系统有两类 。一类是针对可以预料的问题,事先做信息抽取,然后 index 到库里去 支持问答。这类 recall 好,精度也高,但是没有 real time search 的灵活性和以不变应万变。 洪 : 文本信息抽取和理解,全靠 nlp 另一类问答系统就是对通用搜索的直接延伸。利用关键词索引先过滤,把搜罗来的相关网页,在线 parse ,深度分析后找到答案。这个路子技术上是可行的。应对所谓 factoid 问题:何时、何地、谁这样的问题是有效的。(但是复杂问题 如 how 、 why ,还是要走第一类的路线。)为什么可行?因为我们的深度 parsing 是 linear 的效率,在线 parsing 在现代的硬件条件下根本不是问题,瓶颈不在 parsing ,无论多 deep ,比起相关接口之间的延误, parsing 其实是小头。 总之,技术上可以做到立等可取。 对于常见的问题,互联网在线问答系统的 recall 较差根本就不是问题,这是因为网上的冗余信息太多。无论多不堪的 recall ,也不是问题。比如,问 2014 年诺贝尔物理奖得主是谁。这类问题,网上有上百万个答案在。如果关键词过滤了一个子集,里面有几十万答案,少了一个量级,也没问题 。假设在线 nlp 只捞到了其中的十分之一 ,又少了一个量级,那还有几万个instances,这足以满足统计的要求,来坐实 NLP 得来的答案,可以弥补精度上可能的偏差(假设精度有十个百分点的误差)。 IBM 机器在智力竞赛上 beat 人 , 听上去很神奇 , 里面会有很多细节的因应之道,但从宏观上看,一点也不神奇。因为那些个竞赛问题,大多属于 factoid 问题,人受到记忆力有限的挑战,肯定玩不过机器。 雷 : @wei 为什么说事先对材料进行 deep parsing 的搜索不灵活? 事先(pre-parsing)更好。我是主张建立一个超级句法树的。但急于成事的工程师不大愿意。在线做的好处是,内容源可以动态决定。 雷 : 假设一下,我们把谷歌拥有的材料通通进行了 deep parsing ,那么这个搜索会是什么样的? 再辅佐以人工的高级加工 nlp parsing 比关键词索引还是 costs 太大。 雷 : 是,但是现在硬件的条件下,还是可行的吧?那就是把信息转化为了 fact 的知识 是的,哪怕只是把 Google 网页里面的百分之一 parse 一遍 那也有不得了的威力。那是核武器。就是 Powerset Ron 他们当年绘制的图景。 可是 这种大规模运用 NLP 不是我们可定的,成本是一个大因素,还有就是观念和 眼光 , 那是 norvig 这样的人,或其上司才能拍板的。 雷 : 暂时局限在一个领域呢 ? Nick : 可以先小规模吗,如 wiki 等? 雷 : 破坏 google 的力量是 semantic web . 如果每个网站使用的是 semantic web , who needs google, 但是现在的问题是把一个 web2.0 的 site 转化为 web3.0 的成本 Wiki 已经可行,但市场切入点呢? W iki 其实是小菜,比起 我们目前应对的 social media , 无论是量 , 还是语言的难度。 Nick: 但 wiki 有结构 做 wiki 技术上没有任何问题。问题在产品和 businesd model. Nick: 做一个 wiki 的语法树,再叠加 wiki 的结构,已经很有用了。 雷 : wiki 到 dbpedia 还是只有很低的 percentage 吧? Ron 当年 游说你们和微软,不就是 wiki 么,其实他们的 demo ,纯粹 从技术的角度完全可以通过 due diligence 。 大家都知道知识挖掘,在大数据时代有巨大潜力,这是宏观上的认识,永远正确。微观层面,还是要有人在知识基础上做出可挣钱的产品来。微软买P owerset 的时候,肯定也是基于这种宏观认识。但没有后续的产品化,买来的技术就是个负担。 RW: Google 是靠 se 抓流量,然后 ads 赚钱 , Se 技术本身不变现 Nick: @wei powerset 我看过, not impressive at all 那是因为 , 你的角度不同。他们没有把那种结构的威力,用通俗的方式,做成投资人容易看懂的形式。我也玩过 Powerset ,它 的核心能力,其实是有展现的。不过要绕几道弯,才能发现和体会。方向上他们没错。 当然我不是为 Ron 唱赞歌,他再牛,再有名气,他的 parser 比我的还是差远了。这个世界上 yours truly 是第三 -- 如果上帝是第一,在下的下一个系统是第二的话。 当然吹这种牛是得罪人的,不妨当笑话看。 雷 : 呵呵,不用上税,无妨的 Nick: 你的不好意思不得罪人 Jobs不是说过,只有疯狂到以为自己可以改变世界的,才能在雪地里撒尿, 并 留下一些味道或痕迹。 我们是毛时代生人, 自小 有一种精英意识。天将降大任于斯人也,自己吃不饱,也要胸怀世界,解放全人类。老子天下第一的心态就是那种 legacy 。 洪 : Chris Manning 前两年就跟 database/information retrieval 的辩论说,别啥啥 fact db 和 information extraction ,直接 deep parsing 齐活。 雷 : @ 洪 我农民,东西放哪里啊 Parsing real time 的应用场景,东西放内存就可以了,用完就扔,用时再来,现炒现卖。当然那个做不了真正意义上的text mining,只见树木,难见森林。但可以应对搜索引擎对付不了的简单问题。 毛 : 哇哈,不得了。改不改变世界且不说,我的作息时间先被改变了。 雷 : 我以为做机器学习的人在在豪气冲天,原来 @wei 也是! 刘 : @ 雷 一个爱在雪地 …… 洪 : @ 雷 Chris Manning 的意思是, all information are in deep parsed text 雷 : facts 不就是来源于 deep parsed text 吗 洪 : facts are usually triples extracted from text with consensus 。 雷 : under a set of ontologies , these facts form a network , that is, linguistic factors are removed 。 洪 : db ir people dont really believe nlp is a must path for retrieval tasks 雷 : you are right. This is why wei made such big efforts here to point out the problems of those guys. 洪 : linguistic info is transparent to native human speaker , but I don't think it's transparent to computer. So, I believe in communicating with machine, or communicating with people through computer, simpler language in query or logic form should be better. Why do we want to make computer understand human language? It doesn'tmake sense at all. 洪爷说的是哪国话 ? 本来就不存在机器理解语言 , 那个 NLU 只是一个比喻。其实也不存在人工智能,那也是个比喻。 洪 : 现在大多数人可不把 ai/nlu 当比喻 所谓机器理解语言 不过是我们模拟分解了分析理解的过程达到某种表达 representations , 这种表达是达到最终任务的一个方便的桥梁,如此而已。 洪 : 按你的说法,机器人过不了 turing test 这一关 我是回应你为什么要让机器理解语言 。 回答是 , 从来就不需要它去理解。而是因为在人类从形式到内容的映射过程中,我们找到一些路径 , 沿着这个路径我们对人类的理解 , 似乎有一个说得过去的解释。 当然,那位 IR仁兄说的其实是一个具体环节, 指的是搜索框,他说好好的搜索框,给几个关键词就可以查询,既快又好又简单,为什么要把搜索框变成一个自然语言接口,像以前的AskJeeves那样,让人用自然语言提问,然后逼迫机器去理解?从他的角度,这完全不make sense,这种感觉不无道理。明明不用自然语言,多数搜索任务都可以完成得很好,没有道理硬要与机器说“人话”,增加overhead, 还有机器理解过程中的误差。关键词蛮好。互联网搜索这么多年,我们用户其实也被培养出来了,也都习惯了用尽可能少的关键词,以及怎样用关键词的不同组合,容易找到较理想的结果。自然语言接口似乎没有出场的必要。 可是,这只是问题的一个方面。问题是关键词搜索也许可以解决 80% 乃至 90% 的基本信息需求(只是基本,因为心中问题的答案还是需要人在搜索结果中去parse确定,这个过程不总是容易轻松的)。但还有相当一部分问题,我们或者难以用关键词找到线索,或者找出来的所谓相关网页需要太多的人肉阅读还不能搞定。这时候,我们可能就会想,要是有个懂人话的机器,自动解答我们的信息问题多好啊。自然语言接口终究会以某种形式重回台面,增强而不是取代关键词的接口。 洪 : 理解就是 1. 能在人与人之间当二传手; 2. 能根据自己存储的知识和具备的行动能力做出人所认可的反应 说白了 , 就是从线性的言语形式到语法树的映射。这是人类迄今最伟大的发现,或发明,或理论 , 属于最高天机。人类还没有更好的 理论 来解释这个理解过程。这个建树的过程,赶巧可以程序化来模拟,于是诞生了 NLU 毛 : 在图灵测试中,我们是把机器看成黑盒子。但是要让机器通过图灵测试,它就得理解人的语言才能作出反应。 两位大侠,能否推荐几本书看看?最好是科普类的,看着不吃力。 洪爷,不能因为在某些语言任务上,没有语言分析,也做到了,就来否定语言分析的核武器性质。 LSA根本就没有语言分析,但它用到给中学生自动评判作文方面,效果也不错。 洪 : 最近重读了几本认知方面的旧书,我倾向于认为人的内部表征是一种 imaginary 的多维图式表征, linguistic system 只是个人际交流的接口。把多维信息压到线性 。 让计算机理解小说诗歌,估计永远做不到,因为计算机没有人那么强大的 imaginary 内部表征。 @ 毛 wei 和我一起来推荐几本 nlp 方面的书,就像 PDP 一样经典 雷 : @wei 句子的语意理解后的表征方式是什么?还是 tree 吗 ? 逻辑语义 , 这是董老师的表述。外面叫 logical form , 这是从乔老爷那里借来的术语。具体表现细节没必要相同。 雷 : 那么我们把句子给理解后, tree 与 logical form 并存在记忆中? 二者等价。细分可以有:句法树;语义树;语用树。所谓信息抽取 , 就是建语用树。 句法树到语义树 , 就是乔老爷的表层结构到深层结构的逆向转换。 洪 : Chomsky 之所以不谈语义啥的,因为实在没啥科学证据。现在我们所讲的语义都不是 native 的,都是人类的数学逻辑发明,在计算机上热起来的。出口转内销 雷 : 是不是与那时的行为主义为主流有关,因为语意很难有操作定义? 这个讨论越来越高大上 , 也越来越形而上。 毛 : 是啊,再往上一点,就到哲学、认识论的层面了。 另, 跟 PDP 一样经典的是什么书? 乔老爷 57 年小册子。 毛 : 什么书名?我以前只是从编译的角度了解他在形式语言方面的理论(现在也忘了),却不知道他在自然语言方面的贡献。以前我对自然语言毫不关心,也就是这一阵听你们高论才觉得这东西挺有意思。 洪 : 有关语言学和认知科学的科普书, Steven Pinker 写的系列都不错 The Language Instinct (1994) ISBN 978-0-06-097651-4 How the Mind Works (1997) ISBN 978-0-393-31848-7 Words and Rules: The Ingredients of Language (1999) ISBN978-0-465-07269-9 The Blank Slate: The Modern Denial of Human Nature (2002) ISBN978-0-670-03151-1 The Stuff of Thought: Language as a Window into Human Nature(2007) ISBN978-0-670-06327-7 有关 NLP : Dan Jurafsky and James Martin's Speech and Language Processing. 有关基于统计方法的 NLP : Chris Manning and Hinrich Schütze's Foundations of Statistical NaturalLanguage Processing 好像这两本书国内都有影印本 白: 总结一下: wei 的中心意思, nlp 技术在他手里已经很过关了,只是苦于木有好的商业模式,再加上微软谷歌等传统势力的封杀,商业上还不能成大气候。有人建议说回国发展。 deep nlp ,性能不是问题,可以保证线性 online parse ,最坏情形回退到搜索。瓶颈在别处。 雷 : 元芳你怎么看 元芳呢 ? 谢谢白老师的总结 , 实际上就是这么回事。决定成败的不是技术,而是产品方向。技术差,可以砸了产品;技术好,不能保证产品在市场的成功。技术增加的是产品的门槛。 白 : 好的商业模式有两个特点,一个是技术壁垒,一个是侵略性。 nlp 前者不是问题,问题在后者。需要一张极富侵略性的皮。讯飞也有马失前蹄啊。 独 : 多讨论,应该能够找到好的方向。讯飞很多年都做得苦逼死了,熬到这两年才爽。现在做一个新的搜索引擎公司不现实。问答类概念已经被用滥了。出门问问也是因为问答不好做,改作智能手表,反而卖的不错。智能家居的语音交互界面,本质上是一个问答系统。 对于关键词,语法树就是颠覆。 白 : 信息服务三个阶段:门户网站,域名成为商品;搜索引擎,关键词成为商品;社交网络,粉丝成为商品。下一个成为商品的是啥? 问答只是表象,关键是要回答什么成为商品。 分析树也不直接是商品。 白老师说的极是。关键是什么是商品,可以来钱,这个确定了,作为后台的技术产品才有门槛,核武器才能发挥威力。 白 : 我们还是想想,高精准度的 deep nlp 服务,把什么作为标的商品,才能具有侵略性。 Philip: 给 @wei 的高大上技术找个商业模式 独 : 我个人算是比较擅长于设计商业模式的,但是对于 NLP 的直接应用,还是觉得太偏后端,很难找出一个前端产品,对于用户是可感知的刚需。 白 : 不在多而在狠, uber 就够狠。 【相关博文】 【立委科普:自然语言parsers是揭示语言奥秘的LIGO式探测仪】 泥沙龙笔记:创新,失败,再创新,再失败,直至看上去没失败 2015-07-05 泥沙龙笔记:从 sparse data 再论parsing乃是NLP应用的核武器 【置顶:立委科学网博客NLP博文一览(定期更新版)】
【置顶:立委科学网博客NLP博文一览(定期更新版)】 2008-09-20 我们教机器理解语言(Natural Language Understanding),基本的一条就是通过句法分析 (parsing) 解析出句子的意义。什么是一个句子的意义呢?举个例子: John loves Mary. 上述句子有三个构句成分:约翰,玛丽,爱。认识这些词不难,一部词典就成,但这不等于能听懂这个句子,因为句子的意义不是其成分词汇意义的简单相加。同样的词汇,不同的组合,构成不同的句子,就有不同的意义,说明了句子结构分析对于语言理解具有决定性的作用。比较下列各组句子: 1a. John loves Mary. 1b. Mary is loved by John. 1c. John’s love for Mary (is amazing) 1d. Mary’s love by John (is amazing). 2a. Mary loves John. 2b. John is loved by Mary. 2c. Mary’s love for John (is amazing) 2d. John’s love by Mary (is amazing) 3a. John’s Mary is loved. 3b. the love for John’s Mary (is amazing) 4a. Mary’s John is loved. 4b. the love for Mary’s John (is amazing). 以上各组句子里面,虚词和词缀(如is,-ed,’s,the)有所不同,词序排列不同,而基本实词成分是相同的。句式各不相同,有主动态句型,有被动态句型,有用动词love,也有名词love,但是每组的句子中心意义是相同的。句法分析(parsing)的最终目的就是把语言中意义相同但说法不同的句式解码成相同的表达形式(称为逻辑形式 logical form),达成理解。以上述4组句子为例,怎么才叫理解了这些句子呢?如果解析出下列逻辑关系,就可以认为理解了。 1组:约翰是“爱”的的施予者,玛丽是“爱”的对象。 2组:玛丽是“爱”的的施予者,约翰是“爱”的对象。 3组:(约翰的)玛丽是(某人)“爱”的对象。 4组:(玛丽的)约翰是(某人)“爱”的对象。 我们自然语言工作者编制机器语法,为的就是教会机器自动分析(parse)句子,把语言不同句式的种种说法(所谓表层结构 surface structures)解码成如上例所示的能表达结构意义的逻辑关系(所谓深层结构 deep structure)。其重点就是解析动作行为(love)及其施(如约翰)受(如玛丽)关系,即,逻辑主谓宾(logical subject-verb-object SVO)的解构。上述4组句子解构后的形式表达如下: 1组:LOVE: Subj=JOHN; Obj=MARY. 2组:LOVE: Subj=MARY; Obj=JOHN. 3组:LOVE: Obj= 4组:LOVE: Obj= 除了主谓宾的主干以外,句子的意义当然还包括枝节意义,譬如实体的修饰语(e.g. the “beautiful” Mary),行为动作的时间地点条件方式等状语 (e.g. John loves Mary “dearly”),但是,逻辑主谓宾总是句子意义的核心。严格地说,句子的主干应该是“主谓宾补” (S-V-O-C) 四项,因为有些行为动作还需要第二个宾语或者宾语补足语意义才完整。 教会电脑自动理解句子意义有什么用处呢?用处大得很,用处之一是使搜索智能化,直接得到你想要的答案,而不像牵狗一样搜索的结果是成千上万个网页。比如,你有一个疑问:微软收购了哪些公司?你只要告诉带有语言智能的搜索器,Subj=Microsoft, Verb=acquire/buy, Obj=? 逻辑主谓宾武装起来的智能搜索就可以轻易搜得所有媒体报道过的微软兼并过的公司,给你列出一长列来。这是传统搜索引擎 Google, Yahoo, 和 MSN 无法做到的。 下面是笔者开发的英语自动分析机的一个运行实例。输入是英语句子,输出是逻辑主谓宾补。笔者用汉语简单加了一些注解。 这是输入: A U.N. cease-fire resolution has authorized up to 15,000 U.N. peacekeepers to help an equal number of Lebanese troops extend their authority into south Lebanon as Israel withdraws its soldiers. 这是 S-V-O-C 输出: name=”SubjPred” has authorized 动词 A U.N. cease-fire resolution 主语 name=”PredObj” has authorized 动词 up to 15,000 U.N. peacekeepers 宾语 name=”PredInf” has authorized 动词 to help 补语 name=”LSubjPred” to help 动词 up to 15,000 U.N. peacekeepers 主语 name=”PredObj” to help 动词 an equal number of Lebanese troops 宾语 name=”PredComp” to help 动词 extend 补语 name=”LSubjPred” extend 动词 an equal number of Lebanese troops 主语 name=”PredObj” extend 动词 their authority 宾语 name=”PredPrep” extend 动词 into south Lebanon 补语 name=”SubjPred” withdraws 动词 Israel 主语 name=”PredObj” withdraws 动词 its soldiers 宾语 笔者的目标就是制造一台世界上最善解人意的智能机器,大家说的鬼子话它大多听得懂。教机器学人话是既刺激好玩又具有实用价值的干活,笔者教了十几年了,乐此不疲。 Comments (2) yechq 12月 6th, 2008 at 11:18 am edit “笔者的目标就是制造一台世界上最善解人意的智能机器,大家说的鬼子话它大多听得懂。” 好大口气,目前成果如何? liwei 12月 6th, 2008 at 2:19 pm edit 原来是关门吹牛的帖子,出来见光时忘记删改了,不能当真的。 关门在老友中间吹牛基本上与夜行怕鬼吹口哨壮胆类似。呵呵。 谢谢,我去修改一下。