上图第二句的分析，就是我以前说过的结构歧义的应对。你看整个句法树，有三个 O （宾语） 的路径。其中两个是正确的【到-上海】； 【买-上海的飞机票】。第三个 O 【到-上海的飞机票】 是不对的。可以说，“到上海”和“买飞机票”，但不可以说“到飞机票”。这类结构歧义在汉语特别普遍，因为汉语没有宾格，加上汉语的小词 “的”的辖域是一个很大困扰。
QUOTE 甜甜自记事起，就住在这里，水牛城自然是她心目中不可替代的唯一故乡。记得四年前第一次带甜甜回北京探亲，第一天的晚上住在姥姥家，一切对她是那么陌生，没有她已经习惯的美国卡通电视，她满脸委屈地吵着闹着要回家（“I want to go home!”）－－当然是回水牛城的家。我告诉她这就是家呀，是妈妈的家，她怎么也无法认同。
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.
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：
顾: 而且似乎某些高能人群倾向于省略小词。例如华尔街投行和硅谷人士的某些交流中，如果小词太多反而被鄙视，被认为不简洁不性感，这大概是人性，不是中国独有。举一例，出自Liar's Poker, 某trader跳槽，老板以忠诚挽留，他回答，“You want loyalty, hire a cocker spaniel”
顾: long time no see据认为是汉语入侵英语之后产生的，只是大家觉得自然，英美人也用了。这个语句困扰我很久，在网上查了据说是如此，但未必是严肃考证。
我: long time no see 是最直接的展示我东方躶体美女的一个案例。西人突然悟过来，原来语言可以如此简洁，这样地不遮不掩啊。他们觉得可以接受，是因为赶巧这对应了一个常用的语用（pragmatic）场景，朋友见面时候的套话之一，不分中外。在有语用的帮助下，句法可以马虎一些，这也是这类新成语（熟语）形成的背后理由。
说到含金量，其实很多课题，特别是面向应用的课题，并不是什么高精尖的火箭技术（not rocket science），不可能要求一个申请预示某种突破。撰写申请的人是游说方，有责任 highlight 自己的提议里面的亮点，谈方案远景的时候少不了这个突破那个革命的说辞，多少迎合了政府主管部门好大喜功的心态，但实际上很少有多少研究项目会包含那么多闪光的思想和科学研究的革命性转变。（纯科学的研究，突破也不多吧，更何况应用型研究。）应用领域“奇迹”的发生往往植根于细节的积累（所谓 the Devil is in the details），而不是原理上的突破。而对于问题领域的细节，我是有把握的。这是我的长处，也是我提出科研方案比较让人信服的原因。有的时候，不得不有迎合“时尚”的考量，譬如领域里正流行 bootstrapping 等机器自学习的算法，虽然很不成熟，难以解决实际问题，但是基金报告列上它对申请的批准是有益的。不用担心所提议的听上去时尚的方案最后不工作，由于科研的探索性质，最终的解决方案完全可以是另一种路子。说直白了就是，挂羊头卖狗肉不是诚实的科研态度，但是羊头狗头都挂上以后再卖狗肉就没有问题。绝不可以一棵树上吊死。
当然。次范畴就是小规则，小规则优先于大规则。语言规则中，大类的规则（POS-based rules）最粗线条，是默认规则，不涉及具体的次范畴（广义的subcat）。subcat based 的其次。sub-subcat 再其次。一路下推，可以到利用直接量（词驱动）的规则，那是最优先最具体的，包括成语和固定搭配。
王伟DL
文章透露着落地的经验（经历）的光泽，不同的人对此文吸收和反射的谱线也会不同。我贪婪地一连看完，很多地方只觉得在理，的确是是是，有些地方也想表己见，却欲辨已忘言。“...指与大语料库对应的 grammar trees 自动形成的 forest，比 PennTree 大好几个量级。",好羡慕这个大块头！大块头有大智慧！
@算文解字:这篇顶级高手对话，充满思想，可以当武林秘籍参悟的文章，竟然没人转。。。强烈推荐啊！
算文解字
依存关系的确更好用//@立委_米拉: (1) 分层是正道。最起码要两层，基本短语层和句法关系层。（2）顺便一提，作为生成结果，短语结构的表达远不如依存关系的表达。短语结构叠床架屋，不好用，也不够逻辑和普世（不适合词序自由的语言）。当然，这后一点是另外的话题了，不仅仅是 CFG vs FSG 之争了。
算文解字
也对，镜老师批评的是用同一层次的规则处理不同层次的现象的"原教旨"CFG生成方法，提出的对策为FST分层处理。而在CFG下用coarse2fine的（分层）策略，也算是殊途同归了。//@沈李斌AI: 没必要排斥CFG。CFG树是生成结果，不是生成步骤。设计好coarse to fine的生成策略，控制每一步的perplexity和recall
顾: 而且似乎某些高能人群倾向于省略小词。例如华尔街投行和硅谷人士的某些交流中，如果小词太多反而被鄙视，被认为不简洁不性感，这大概是人性，不是中国独有。举一例，出自Liar's Poker, 某trader跳槽，老板以忠诚挽留，他回答，“You want loyalty, hire a cocker spaniel”
顾: long time no see据认为是汉语入侵英语之后产生的，只是大家觉得自然，英美人也用了。这个语句困扰我很久，在网上查了据说是如此，但未必是严肃考证。
我: long time no see 是最直接的展示我东方躶体美女的一个案例。西人突然悟过来，原来语言可以如此简洁，这样地不遮不掩啊。他们觉得可以接受，是因为赶巧这对应了一个常用的语用（pragmatic）场景，朋友见面时候的套话之一，不分中外。在有语用的帮助下，句法可以马虎一些，这也是这类新成语（熟语）形成的背后理由。
说到含金量，其实很多课题，特别是面向应用的课题，并不是什么高精尖的火箭技术（not rocket science），不可能要求一个申请预示某种突破。撰写申请的人是游说方，有责任 highlight 自己的提议里面的亮点，谈方案远景的时候少不了这个突破那个革命的说辞，多少迎合了政府主管部门好大喜功的心态，但实际上很少有多少研究项目会包含那么多闪光的思想和科学研究的革命性转变。（纯科学的研究，突破也不多吧，更何况应用型研究。）应用领域“奇迹”的发生往往植根于细节的积累（所谓 the Devil is in the details），而不是原理上的突破。而对于问题领域的细节，我是有把握的。这是我的长处，也是我提出科研方案比较让人信服的原因。有的时候，不得不有迎合“时尚”的考量，譬如领域里正流行 bootstrapping 等机器自学习的算法，虽然很不成熟，难以解决实际问题，但是基金报告列上它对申请的批准是有益的。不用担心所提议的听上去时尚的方案最后不工作，由于科研的探索性质，最终的解决方案完全可以是另一种路子。说直白了就是，挂羊头卖狗肉不是诚实的科研态度，但是羊头狗头都挂上以后再卖狗肉就没有问题。绝不可以一棵树上吊死。
Sydney Brenner, 索尔克生物研究所高级研究员（2002年诺贝尔奖得主，在基因编码领域有突出贡献）
Marvin Minsky, 麻省理工学院媒体艺术与科学教授
Noam Chomsky, 麻省理工学院语言与哲学系教授
Emilio Bizzi, 麻省理工学院脑科学研究所教授
Barbara H. Partee 麻省大学语言与哲学系教授
Patrick H. Winston 麻省理工学院人工智能与计算机科学教授
Chomsky的主要观点：
A. Chomsky认为统计语言模型取得过工程意义上的成功，但不关科学的事。
B. 为语言事实建模就像收集蝴蝶标本。科学（尤其是语言学）想要的是基本原则。
C. 统计模型无法理解，并不是关于研究对象的洞见。
D. 统计模型或许可以对一些现象做出精确的模拟，但这是迷途。人们并不根据前面出现的两个单词去预测后面一个单词。人们生成句子（词语序列）的方式是从内在的语义到树结构，再到表层的线性词语序列。
E. 统计模型已经被证实无法用于学习语言。因此语言必然是天生的。用语言模型去解释语言是浪费时间。
Norvig的主要回应：
A. 工程上的成功确实不是科学目标。不过科学和工程是比翼齐飞的。工程上的成功可以作为科学上成功模型的证据。
B. 科学是事实和理论的混合体。理论过分凌驾于事实之上并不可取。在科学史上，不断积累事实是科研正途，并非异类。关于语言的科学也不应例外。
C. 包含几十亿个参数的统计模型确实难以直观理解。个人确实无法核查每个个体参数的意义所在。但是，人们可以通过了解整个模型的特性而获得对于统计模型合理与否的认知：即一个统计模型是怎样有效的，或者为什么无效，它是如何从数据中学到模型函数的，等等。
D. 基于词概率的Markov（马尔科夫模型）确实无法对所有的语言现象建模。这就像没有概率的简单树结构模型无法对所有的语言现象建模一样。我们需要的语言模型是可以覆盖词、树结构、语义、上下文、语篇等等不同层次语言现象的更复杂的概率模型。Chomsky不能因为旧的统计模型的缺点就一概否定所有的统计语言模型。研究如何解释语言（比如语音识别）的人当中，绝大多数人都认同，解释是一个概率问题。当一个语音流到了我耳朵里，要把这串语音流恢复为说话者的意义，是一个概率问题。爱因斯坦说过，让事情变得简单，直到不能再简单为止。许多科学现象都有随机性。最简单的模型就是概率模型。语言也是这样一种现象。因此概率模型是表达语言事实的最好工具。
E. 1967年，Gold定理指出了形式化的数学语言在逻辑推导上的理论限制。但是，这跟自然语言学习者面临的问题毫无关系。无论如何，在1969年，我们就知道了，概率推理不受这一限制的约束（Horning证明学习概率上下文无关文法PCFG是可能的）。我同意Chomsky所说的，人类具有学习语言的天赋。但是我们对如何获得概率化的语言表示，对统计学习，都还缺乏足够的知识。我认为很可能人类学习语言涉及到概率和统计推理，但是我们并不清楚细节。
1） I never, ever, ever, ever, ... fiddle around in any way with electrical equipment.
2） She never, ever, ever, ever, ... fiddles around in any way with electrical equipment.
3） * I never, ever, ever, ever, ... fiddles around in any way with electrical equipment.
4） * She never, ever, ever, ever, ... fiddle around in any way with electrical equipment.
· "It is neutral green, colorless green, like the glaucous water lying in a cellar." The Paris we remember, Elisabeth Finley Thomas (1942).
· "To specify those green ideas is hardly necessary, but you may observe Mr. [D. H.] Lawrence in the role of the satiated aesthete." The New Republic: Volume 29 p. 184, William White (1922).
· "Ideas sleep in books." Current Opinion: Volume 52, (1912).
· "Not gonna do it. Wouldn't be prudent." (Dana Carvey, impersonating George H. W. Bush)
· "Thinks he can outsmart us, does he?" (Evelyn Waugh, The Loved One)
· "Likes to fight, does he?" (S.M. Stirling, The Sunrise Lands)
· "Thinks he's all that." (Kate Brian, Lucky T)
· "Go for a walk?" (countless dog owners)
· "Gotcha!" "Found it!" "Looks good to me!" (common expressions)
语言学家可以为如何解释上面这些现象争个没完没了。但语言的多样性似乎远比用布尔值（true or false）来描述pro-drop参数值要复杂。一个理论框架不应该把简单性置于反映现实的准确性之上。
技术不是问题（笨蛋不算，你要是找到一个只会忽悠的笨蛋，那是 due diligence 太差，怨不得人）。
Nick: 嗨，老套路，骂别人是为了夸自个。
可不，卖瓜王爷。不过，那也是客观事实，内举不避己，不能因为自己能就偏要说不能，最后还是要系统说话。
当然，这玩意儿要做好（精准达到接近人的分析能力，鲁棒达到可以对付社会媒体这样的monster，高效达到线性实现，real time 应用），确实不是一蹴而就能成的。这里有个n万小时定律。大体是，NLP入门需要一万小时（大约五年工龄），找到感觉需要两万小时，栽几个有意义的跟头需要三万小时，得心应手需要四万小时，等你做到五万小时（入行25年）还没被淘汰的话，就可以成精了。那是一种有如神助、如入无人之境的感觉，体会的人不多。打住。
为了所谓语言的递归性，人脑，或电脑，必须有个堆栈的结构才好，这离语言事实太远，也违背了人脑短期记忆的限制。世界上哪里有人说话，只管开门而不关门，只加左括号不加右括号，一直悬着吊着的？最多三重门吧，一般人就受不了了。就算你是超人，你受得了，你的受众也受不了，无法 parse 啊。说话不是为了交流，难道是故意难为人，为了人不懂你而说话？不 make sense 嘛。
Quora has a question with discussions on "Why is machine learning used heavily for Google's ad ranking and less for their search ranking?" A lot of people I've talked to at Google have told me that the ad ranking system is largely machine learning based, while search ranking is rooted in functions that are written by humans using their intuition (with some components using machine learning).
Surprise? Contrary to what many people have believed, Google search consists of hand-crafted functions using heuristics. Why?
479
One very popular reply there is from Edmond Lau, Ex-Google Search Quality Engineer who said something which we have been experiencing and have indicated over and over in my past blogs on Machine Learning vs. Rule System, i.e. it is very difficult to debug an ML system for specific observed quality bugs while the rule system, if designed modularly, is easy to control for fine-tuning:
From what I gathered while I was there, Amit Singhal, who heads Google's core ranking team, has a philosophical bias against using machine learning in search ranking. My understanding for the two main reasons behind this philosophy is:
In a machine learning system, it's hard to explain and ascertain why a particular search result ranks more highly than another result for a given query. The explainability of a certain decision can be fairly elusive; most machine learning algorithms tend to be black boxes that at best expose weights and models that can only paint a coarse picture of why a certain decision was made.
Even in situations where someone succeeds in identifying the signals that factored into why one result was ranked more highly than other, it's difficult to directly tweak a machine learning-based system to boost the importance of certain signals over others in isolated contexts. The signals and features that feed into a machine learning system tend to only indirectly affect the output through layers of weights, and this lack of direct control means that even if a human can explain why one web page is better than another for a given query, it can be difficult to embed that human intuition into a system based on machine learning.
Rule-based scoring metrics, while still complex, provide a greater opportunity for engineers to directly tweak weights in specific situations. From Google's dominance in web search, it's fairly clear that the decision to optimize for explainability and control over search result rankings has been successful at allowing the team to iterate and improve rapidly on search ranking quality. The team launched 450 improvements in 2008 [1], and the number is likely only growing with time.
Ads ranking, on the other hand, tends to be much more of an optimization problem where the quality of two ads are much harder to compare and intuit than two web page results. Whereas web pages are fairly distinctive and can be compared and rated by human evaluators on their relevance and quality for a given query [2], the short three- or four-line ads that appear in web search all look fairly similar to humans. It might be easy for a human to identify an obviously terrible ad, but it's difficult to compare two reasonable ones:
Branding differences, subtle textual cues, and behavioral traits of the user, which are hard for humans to intuit but easy for machines to identify, become much more important. Moreover, different advertisers have different budgets and different bids, making ad ranking more of a revenue optimization problem than merely a quality optimization problem. Because humans are less able to understand the decision behind an ads ranking decision that may work well empirically, explainability and control -- both of which are important for search ranking -- become comparatively less useful in ads ranking, and machine learning becomes a much more viable option.
Edmond Lau's answer is great, but I wanted to add one more important piece of information.
When I was on the search team at Google (2008-2010), many of the groups in search were moving away from machine learning systems to the rules-based systems. That is to say that Google Search used to use more machine learning, and then went the other direction because the team realized they could make faster improvements to search quality with a rules based system. It's not just a bias, it's something that many sub-teams of search tried out and preferred.
I was the PM for Images, Video, and Local Universal - 3 teams that focus on including the best results when they are images, videos, or places. For each of those teams I could easily understand and remember how the rules worked. I would frequently look at random searches and their results and think "Did we include the right Images for this search? If not, how could we have done better?". And when we asked that question, we were usually able to think of signals that would have helped - try it yourself. The reasons why *you* think we should have shown a certain image are usually things that Google can actually figure out.
Part of the answer is legacy, but a bigger part of the answer is the difference in objectives, scope and customers of the two systems.
The customer for the ad-system is the advertiser (and by proxy, Google's sales dept). If the machine-learning system does a poor job, the advertisers are unhappy and Google makes less money. Relatively speaking, this is tolerable to Google. The system has an objective function ($) and machine learning systems can be used when they can work with an objective function to optimize. The total search-space (# of ads) is also much much smaller.
The search ranking system has a very subjective goal - user happiness. CTR, query volume etc. are very inexact metrics for this goal, especially on the fringes (i.e. query terms that are low-volume/volatile). While much of the decisioning can be automated, there are still lots of decisions that need human intuition.
To tell whether site A better than site B for topic X with limited behavioural data is still a very hard problem. It degenerates into lots of little messy rules and exceptions that tries to impose a fragile structure onto human knowledge, that necessarily needs tweaking.
An interesting question is - is the Google search index (and associated semantic structures) catching up (in size and robustness) to the subset of the corpus of human knowledge that people are interested in and searching for ?
My guess is that right now, the gap is probably growing - i.e. interesting/search-worthy human knowledge is growing faster than Google's index.. Amit Singhal's job is probably getting harder every year. By extension, there are opportunities for new search providers to step into the increasing gap with unique offerings.
p.s: I used to manage an engineering team for a large search provider (many years ago).
A couple of days ago I had coffee with Peter Norvig. Peter is currently Director of Research at Google. For several years until recently, he was the Director of Search Quality -- the key man at Google responsible for the quality of their search results. Peter also is an ACM Fellow and co-author of the best-selling AI textbook Artificial Intelligence: A Modern Approach. As such, Peter's insights into search are truly extraordinary.
I have known Peter since 1996, when he joined a startup called Junglee, which I had started together with some friends from Stanford. Peter was Chief Scientist at Junglee until 1998, when Junglee was acquired by Amazon.com. I've always been a great admirer of Peter and have kept in touch with him through his short stint at NASA and then at Google. He's now taking a short leave of absence from Google to update his AI textbook. We had a fascinating discussion, and I'll be writing a couple of posts on topics we covered.
It has long been known that Google's search algorithm actually works at 2 levels:
An offline phase that extracts "signals" from a massive web crawl and usage data. An example of such a signal is page rank. These computations need to be done offline because they analyze massive amounts of data and are time-consuming. Because these signals are extracted offline, and not in response to user queries, these signals are necessarily query-independent. You can think of them tags on the documents in the index. There are about 200 such signals.
An online phase, in response to a user query. A subset of documents is identified based on the presence of the user's keywords. Then, these documents are ranked by a very fast algorithm that combines the 200 signals in-memory using a proprietary formula.
The online, query-dependent phase appears to be made-to-order for machine learning algorithms. Tons of training data (both from usage and from the armies of "raters" employed by Google), and a manageable number of signals (200) -- these fit the supervised learning paradigm well, bringing into play an array of ML algorithms from simple regression methods toSupport Vector Machines. And indeed, Google has tried methods such as these. Peter tells me that their best machine-learned model is now as good as, and sometimes better than, the hand-tuned formula on the results quality metrics that Google uses.
The big surprise is that Google still uses the manually-crafted formula for its search results. They haven't cut over to the machine learned model yet. Peter suggests two reasons for this. The first is hubris: the human experts who created the algorithm believe they can do better than a machine-learned model. The second reason is more interesting. Google's search team worries that machine-learned models may be susceptible to catastrophic errors on searches that look very different from the training data. They believe the manually crafted model is less susceptible to such catastrophic errors on unforeseen query types.
This raises a fundamental philosophical question. If Google is unwilling to trust machine-learned models for ranking search results, can we ever trust such models for more critical things, such as flying an airplane, driving a car, or algorithmic stock market trading? All machine learning models assume that the situations they encounter in use will be similar to their training data. This, however, exposes them to the well-known problem of induction in logic.
The classic example is the Black Swan, popularized by Nassim Taleb'seponymous book. Before the 17th century, the only swans encountered in the Western world were white. Thus, it was reasonable to conclude that "all swans are white." Of course, when Australia was discovered, so were the black swans living there. Thus, a black swan is a shorthand for something unexpected that is outside the model.
Taleb argues that black swans are more common than commonly assumed in the modern world. He divides phenomena into two classes:
Mediocristan, consisting of phenomena that fit the bell curve model, such as games of chance, height and weight in humans, and so on. Here future observations can be predicted by extrapolating from variations in statistics based on past observation (for example, sample means and standard deviations).
Extremistan, consisting of phenomena that don't fit the bell curve model, such as the search queries, the stock market, the length of wars, and so on. Sometimes such phenomena can sometimes be modeled using power laws or fractal distributions, and sometimes not. In many cases, the very notion of a standard deviation is meaningless.
Taleb makes a convincing case that most real-world phenomena we care about actually inhabit Extremistan rather than Mediocristan. In these cases, you can make quite a fool of yourself by assuming that the future looks like the past.
The current generation of machine learning algorithms can work well in Mediocristan but not in Extremistan. The very metrics these algorithms use, such as precision, recall, and root-mean square error (RMSE), make sense only in Mediocristan. It's easy to fit the observed data and fail catastrophically on unseen data. My hunch is that humans have evolved to use decision-making methods that are less likely blow up on unforeseen events (although not always, as the mortgage crisis shows).
I'll leave it as an exercise to the interested graduate student to figure out whether new machine learning algorithms can be devised that work well in Extremistan, or prove that it cannot be done.
TRACKBACK
TrackBack URL for this entry: http://www.typepad.com/services/trackback/6a00d83471bc3153ef00e5527c00a38833
有问，这一波热潮会不会是类似2000年的又一个巨大的泡沫？我的观察是，也是，也不是。的确，在大数据的市场还不成熟，发展和盈利模式还很不清晰的时候，大家一窝蜂拥上来创业、投资和冒险，其过热的行为模式确实让人联想到世纪之交的互联网 dot com 的泡沫。然而，这次热潮不是泡沫那么简单，里面蕴含了实实在在的内容和价值潜力，我们下面会具体谈到。当然这些潜在价值与市场的消化能力是否匹配，仍是一个巨大的问题。可以预见三五年之后的情景，涅磐的凤凰和死在沙滩上的前浪共同谱写了大数据交响乐的第一乐章。
所谓大数据，更多的是社会媒体火热以后的专指，是已经与施事背景相关联的数据，而不是搜索引擎从开放互联网搜罗来的混杂集合。没有社会媒体及其用户社会网络作为背景，纯粹从量上看，“大数据”早就存在了，它催生了搜索产业。对于搜索引擎，big data 早已不是新的概念，面对互联网的汪洋大海，搜索巨头利用关键词索引（keyword indexing）为亿万用户提供搜索服务已经很多年了。我们每一个网民都是受益者，很难想象一个没有搜索的互联网世界。但那不是如今的 buzz word，如今的大数据与社会媒体密不可分。当然，数据挖掘领域把用户信息和消费习惯的数据结合起来，已经有很多成果和应用。自然语言的大数据可以看作是那个应用的继续，从术语上说就是，文本挖掘（text mining，from social media big data）是数据挖掘（data mining） 的自然延伸。对于语言技术，NLP 系统需要对语言做结构分析，理解其语义，这样的智能型工作比给关键词建立索引要复杂百倍，也因此 big data scale up 一直是自然语言技术的一个瓶颈。
大数据时代只认数据不认人。Of course, In God We Trust. But in everything else we need data. 道理很简单，在信息爆炸的时代，任何个人的精力、能力和阅历都是有限的，所看到听到的都是冰山一角。大V也是如此，大家都在盲人摸象。唯有大数据挖掘才有资格为纵览全貌提供导引。
在处理海量数据的问题解决以后，查准率和查全率变得相对不重要了。换句话说，即便不是最优秀的系统，只有平平的查准率（譬如70%，抓100个，只有70个抓对了），平平的查全率（譬如30%，三个只能抓到一个），只要可以用于大数据，一样可以做出优秀的实用系统来。其根本原因在于两个因素：一是大数据时代的信息冗余度；二是人类信息消化的有限度。查全率的不足可以用增加所处理的数据量来弥补，这一点比较好理解。既然有价值的信息，有统计意义的信息，不可能是“孤本”，它一定是被许多人以许多不同的说法重复着，那么查全率不高的系统总会抓住它也就没有疑问了。从信息消费者的角度，一个信息被抓住一千次，与被抓住900次，是没有本质区别的，信息还是那个信息，只要准确就成。疑问在一个查准率不理想的系统怎么可以取信于用户呢？如果是70%的系统，100条抓到的信息就有30条是错的，这岂不是鱼龙混杂，让人无法辨别，这样的系统还有什么价值？沿着这个思路，别说70%，就是高达90%的系统也还是错误随处可见，不堪应用。这样的视点忽略了实际的挖掘系统中的信息筛选（sampling）与整合（fusion）的环节，因此夸大了系统的个案错误对最终结果的负面影响。实际上，典型的情景是，面对海量信息源，信息搜索者的几乎任何请求，都会有数不清的潜在答案。由于信息消费者是人，不是神，即便有一个完美无误的理想系统能够把所有结果，不分巨细都提供给他，他也无福消受（所谓 information overload）。因此，一个实用系统必须要做筛选整合，把统计上最有意义的结果呈现出来。这个筛选整合的过程是挖掘的一部分，可以保证最终结果的质量远远高于系统的个案质量。总之，size matters，多了就不一样了。大数据改变了技术应用的条件和生态，大数据更能将就不完美的引擎。
大数据不是决策的唯一依据，只是依据之一。正确的决策必须综合各种信息来源。大事不提，看看笔者购买洗衣机是怎样使用大数据、朋友口碑、实地考察以及种种其他考量的吧。以为有了大数据，就万事大吉，是不切实际的。值得注意的是，即便被认为是真实反映的同一组数据结果也完全可能有不同的解读（interpretations），人们就是在这种解读的争辩中逼近真相。一个好的大数据系统，必须创造条件，便于用户 drill down 去验证或否定一种解读，便于用户通过不同的条件限制及其比较来探究真相。
所列“成见”有两类：一类是“偏”见，如【成见一】至【成见五】。这类偏见主要源于不完全归纳，他们也许看到过或者尝试过规则系统某一个类型,浅尝辄止，然后遽下结论（jump to conclusions）。盗亦有道，情有可原，虽然还是应该对其一一纠“正”。成见的另一类是谬见，可以事实证明其荒谬。令人惊诧的是，谬见也可以如此流行。【成见五】以降均属不攻自破的谬见。譬如【成见八】说规则系统只能分析规范性语言。事实胜于雄辩，我们开发的以规则体系为主的舆情挖掘系统处理的就是非规范的社交媒体。这个系统的大规模运行和使用也驳斥了【成见六】。
米拉宝鉴：确实应该展开讨论，不着急，慢慢来。所罗列的“偏见”有两类：一类是谬见，可以证明其荒谬，譬如说规则系统不能处理社会媒体，只能分析规范性语言。另一类就是“偏”见，盗亦有道，情有可原，虽然还是应该对其纠“正”。这类偏见主要源于不完全归纳，他们也许看到过或者尝试过规则系统某一个类型。 浅尝辄止，然后 jump to conclusion
回复 ： 改弦易辙没有问题。从一个 school 转学到一个新 school 很自然，我要是年轻20岁，也一定加入 converting 的潮流。本文揭示的是偏见为什么如此流行，被很多高智商学者视为理所当然，乃至于不得不怀疑宗教疑似的世界观在作祟。至于翘翘板现象，又称按下葫芦起了瓢的问题，以后单论，其实是有有效对策的。当然，也必须承认统计路线的性质决定了它们比较善于在多种因素中玩平衡。
转载
You can follow this conversation by subscribing to the comment feed for this post.
，也许是我的电脑有问题
