On Chomsky and The Two Cultures of Statistical Learning

On Chomsky and the Two Cultures of Statistical Learning
11-6-1 9:43

At the Brains, Minds, and Machines symposium held during MIT's 150th birthday party, Technology Review reports that Prof. Noam Chomsky derided researchers in machine learning who use purely statistical methods to produce behavior that mimics something in the world, but who don't try to understand the meaning of that behavior. Chomsky compared such researchers to scientists who might study the dance made by a bee returning to the hive, and who could produce a statistically based simulation of such a dance without attempting to understand why the bee behaved that way. "That's a notion of [scientific] success that's very novel. I don't know of anything like it in the history of science," said Chomsky.
MIT: 150
This essay discusses what Chomsky said, speculates on what he might have meant, and tries to determine the truth and importance of his claims. As I recall, Chomsky's remarks were in response to an audience member who asked about the success of statistical models for language tasks such as speech recognition and machine translation. Chomsky's reply (not recorded by Technology Review) was something along the lines that such models had some limited success in some application areas, but not the kind of success that counted. So the questions are: 1. 2. 3. 4. 5. What did Chomsky mean, and is he right? What is a statistical model? How successful are statistical language models? Is there anything like their notion of success in the history of science? What doesn't Chomsky like about statistical models?
Noam Chomsky
What did Chomsky mean, and is he right?

I take Chomsky's points to be the following: A. Statistical language models have had engineering success, but that is irrelevant to science. B. Accurately modeling linguistic facts is just butterfly collecting; what matters in science (and specifically linguistics) is the underlying principles. C. Statistical models are incomprehensible; they provide no insight. D. Statistical models may provide an accurate simulation of some phenomena, but the simulation is done completely the wrong way; people don't decide what the third word of a sentence should be by consulting a probability table keyed on the previous two words, rather they map from an internal semantic form to a syntactic tree-structure, which is then linearized into words. This is done without any probability or statistics. E. Statistical models have been proven incapable of learning language; therefore language must be
http://norvig.com/chomsky.html 1 29
11-6-1 9:43
innate, so why are these statistical modelers wasting their time on the wrong enterprise? Is he right? That's a long-standing debate. These are my answers: A. I agree that engineering success is not the goal or the measure of science. But I observe that science and engineering develop together, and that engineering success shows that something is working right, and so is evidence (but not proof) of a scientifically successful model. B. Science is a combination of gathering facts and making theories; neither can progress on its own. I think Chomsky is wrong to push the needle so far towards theory over facts; in the history of science, the laborious accumulation of facts is the dominant mode, not a novelty. The science of understanding language is no different than other sciences in this respect. C. I agree that it can be difficult to make sense of a model containing billions of parameters. Certainly a human can't understand such a model by inspecting the values of each parameter individually. But one can gain insight by examing the properties of the modelwhere it succeeds and fails, how well it learns as a function of data, etc. D. I agree that a Markov model of word probabilities cannot model all of language. It is equally true that a concise tree-structure model without probabilities cannot model of all language. What is needed is a probabilistic model that covers words, trees, semantics, context, discourse, etc. Chomsky dismisses all probabilistic models because of shortcomings of particular 50-year old models. I understand how Chomsky arrives at the conclusion that probabilistic models are unnecessary, from his study of the generation of language. But the vast majority of people who study interpretation tasks, such as speech recognition, quickly see that interpretation is an inherently probabilistic problem: given a stream of noisy input to my ears, what did the speaker most likely mean? Einstein said to make everything as simple as possible, but no simpler. Many phenomena in science are stochastic, and the simplest model of them is a probabilistic model; I believe language is such a phenomenon and therefore that probabilistic models are our best tool for representing facts about language, for algorithmically processing language, and for understanding how humans process language. E. In 1967, Gold's Theorem showed some theoretical limitations of logical deduction on formal mathematical languages. But this result has nothing to do with the task faced by learners of natural language. In any event, by 1969 we knew that probabilistic inference (over probabilistic context-free grammars) is not subject to those limitations (Horning showed that learning of PCFGs is possible). I agree with Chomsky that it is undeniable that humans have some innate capability to learn natural language, but we don't know enough about that capability to rule out probabilistic language representations, nor statistical learning. I think it is much more likely that human language learning involves something like probabilistic and statistical inference, but we just don't know yet. Now let me back up my answers with a more detailed look at the remaining questions.
What is a statistical model?

A statistical model is a mathematical model which is modified or trained by the input of data points. Statistical models are often but not always probabilistic. Where the distinction is important we will be careful not to just say "statistical" but to use the following component terms: A mathematical model specifies a relation among variables, either in functional form that maps
11-6-1 9:43
inputs to outputs (e.g. y = m x + b) or in relation form (e.g. the following (x, y) pairs are part of the relation). A probabilistic model specifies a probability distribution over possible values of random variables, e.g., P(x, y), rather than a strict deterministic relationship, e.g., y = f(x). A trained model uses some training/learning algorithm to take as input a collection of possible models and a collection of data points (e.g. (x, y) pairs) and select the best model. Often this is in the form of choosing the values of parameters (such as m and b above) through a process of statistical inference. For example, a decade before Chomsky, Claude Shannon proposed probabilistic models of communication based on Markov chains of words. If you have a vocabulary of 100,000 words and a second-order Markov model in which the probability of a word depends on the previous two words, then you need a quadrillion (10 15) probability values to specify the model. The only feasible way to learn these 1015 values is to gather statistics from data and introduce some smoothing method for the many cases where there is no data. Therefore, most (but not all) probabilistic models are trained. Also, many (but not all) trained models are probabilistic.
Claude Shannon
As another example, consider the Newtonian model of gravitational attraction, which says that the force between two objects of mass m1 and m2 a distance r apart is given by F = G m1 m2 / r 2 where G is the universal gravitational constant. This is a trained model because the gravitational constant G is determined by statistical inference over the results of a series of experiments that contain stochastic experimental error. It is also a deterministic (non-probabilistic) model because it states an exact functional relationship. I believe that Chomsky has no objection to this kind of statistical model. Rather, he seems to reserve his criticism for statistical models like Shannon's that have quadrillions of parameters, not just one or two. (This example brings up another distinction: the gravitational model is continuous and quantitative whereas the linguistic tradition has favored models that are discrete, categorical, and qualitative: a word is or is not a verb, there is no question of its degree of verbiness. For more on these distinctions, see Chris Manning's article on Probabilistic Syntax.) A relevant probabilistic statistical model is the ideal gas law, which describes the pressure P of a gas in terms of the the number of molecules, N, the temperature T, and Boltzmann's constant, K: P = N k T / V. The equation can be derived from first principles using the tools of statistical mechanics. It is an uncertain, incorrect model; the true model would have to describe the motions of individual gas molecules. This model ignores that complexity and summarizes our uncertainty about the location of individual molecules. Thus, even though it is statistical and probabilistic, even though it does not completely model reality, it does provide both good predictions and insightinsight that is not available from trying to understand the true movements of individual molecules.
11-6-1 9:43
Now let's consider the non-statistical model of spelling expressed by the rule "I before E except after C." Compare that to the probabilistic, trained statistical model:
P(IE) = 0.0177 P(EI) = 0.0046 P(CIE) = 0.0014 P(CEI) = 0.0005
This model comes from statistics on a corpus of a trillion words of English text. The notation P(IE) is the probability that a word sampled from this corpus contains the consecutive letters "IE." The statistical data confirms that IE is in fact about 3.8 times more common than EI, but it refutes the "except after C" part; in fact, "CIE" is more common than "CEI" (although only 2.8 times). Examples of "CIE" words include "science," "society," "ancient" and "species." The disadvantage of the "I before E except after C" model is that it is inaccurate, making the wrong prediction 20% of the time. A fullystatistical model (say, one that gave the probability of all 4-letter sequences, and/or of all known words) could be ten times more accurate at the task of spelling, but offers little insight into what is going on. (Insight would require a model that knows about phonemes, syllabification, and language of origin. Such a model could be trained (or not) and probabilistic (or not).) As a final example (not of statistical models, but of insight), consider the Theory of Supreme Court Justice Hand-Shaking: when the supreme court convenes, all attending justices shake hands with every other justice. The number of attendees, n, must be an integer in the range 0 to 9; what is the total number of handshakes, h for a given n? Here are three possible explanations: A. Each of n justices shakes hands with the other n - 1 justices, but that counts Alito/Breyer and Breyer/Alito as two separate shakes, so we should cut the total in half, and we end up with h = n (n - 1) / 2. B. To avoid double-counting, we will order the justices by seniority and only count a moresenior/more-junior handshake, not a more-junior/more-senior one. So we count, for each justice, the shakes with the more junior justices, and sum them up, giving h = i = 1 .. n (i - 1). C. Just look at this table: n: 0 1 2 3 4 5 6 7 8 9 h: 0 0 1 3 6 10 15 21 28 36 Some people might prefer A, some might prefer B, and if you are slow at doing multiplication or addition you might prefer C. Why? All three explanations describe exactly the same theory the same function from n to h, over the entire domain of possible values of n. Thus we could prefer A (or B) over C only for reasons other than the theory itself. We might find that A or B gave us a better understanding of the problem. A and B are certainly more useful than C for figuring out what happens if Congress exercises its power to add an additional associate justice. Theory A might be most helpful in developing a theory of handshakes at the end of a hockey game (when each player shakes hands with players on the opposing team) or in proving that the number of people who shook an odd number of hands at the MIT Symposium is even.
How successful are statistical language models?

Chomsky said words to the effect that statistical language models have had some limited success in some application areas. Let's look at computer systems that deal with language, and at the notion of
11-6-1 9:43
"success" defined by "making accurate predictions about the world." First, the major application areas: Search engines: 100% of major players are trained and probabilistic. Their operation cannot be described by a simple function. Speech recognition: 100% of major systems are trained and probabilistic, mostly relying on probabilistic hidden Markov models. Machine translation: 100% of top competitors in competitions such as NIST use statistical methods. Some commercial systems use a hybrid of trained and rule-based approaches. Of the 4000 language pairs covered by machine translation systems, a statistical system is by far the best for every pair except Japanese-English, where the top statistical system is roughly equal to the top hybrid system. Question answering: this application is less well-developed, and many systems build heavily on the statistical and probabilistic approach used by search engines. The IBM Watson system that recently won on Jeopardy is thoroughly probabilistic and trained, while Boris Katz's START is a hybrid. All systems use at least some statistical techniques. Now let's look at some components that are of interest only to the computational linguist, not to the end user: Word sense disambiguation: 100% of top competitors at the SemEval-2 competition used statistical techniques; most are probabilistic; some use a hybrid approach incorporating rules from sources such as Wordnet. Coreference resolution: The majority of current systems are statistical, although we should mention the system of Haghighi and Klein, which can be described as a hybrid system that is mostly rule-based rather than trained, and performs on par with top statistical systems. Part of speech tagging: Most current systems are statistical. The Brill tagger stands out as a successful hybrid system: it learns a set of deterministic rules from statistical data. Parsing: There are many parsing systems, using multiple approaches. Almost all of the most successful are statistical, and the majority are probabilistic (with a substantial minority of deterministic parsers). Clearly, it is inaccurate to say that statistical models (and probabilistic models) have achieved limited success; rather they have achieved a dominant (although not exclusive) position. Another measure of success is the degree to which an idea captures a community of researchers. As Steve Abney wrote in 1996, "In the space of the last ten years, statistical methods have gone from being virtually unknown in computational linguistics to being a fundamental given. ... anyone who cannot at least use the terminology persuasively risks being mistaken for kitchen help at the ACL [Association for Computational Linguistics] banquet." Now of course, the majority doesn't rule -- just because everyone is jumping on some bandwagon, that doesn't make it right. But I made the switch: after about 14 years of trying to get language models to work using logical rules, I started to adopt probabilistic approaches (thanks to pioneers like Gene Charniak (and Judea Pearl for probability in general) and to my colleagues who were early adopters, like Dekai Wu). And I saw everyone around me making the same switch. (And I didn't see anyone going in the other direction.) We all saw the limitations of the old tools, and the benefits of the new. And while it may seem crass and anti-intellectual to consider a financial measure of success, it is worth
11-6-1 9:43
noting that the intellectual offspring of Shannon's theory create several trillion dollars of revenue each year, while the offspring of Chomsky's theories generate well under a billion. This section has shown that one reason why the vast majority of researchers in computational linguistics use statistical models is an engineering reason: statistical models have state-of-the-art performance, and in most cases non-statistical models perform worst. For the remainder of this essay we will concentrate on scientific reasons: that probabilistic models better represent linguistic facts, and statistical techniques make it easier for us to make sense of those facts.
Is there anything like [the statistical model] notion of success in the history of science?
When Chomsky said "That's a notion of [scientific] success that's very novel. I don't know of anything like it in the history of science" he apparently meant that the notion of success of "accurately modeling the world" is novel, and that the only true measure of success in the history of science is "providing insight" of answering why things are the way they are, not just describing how they are. A dictionary definition of science is "the systematic study of the structure and behavior of the physical and natural world through observation and experiment," which stresses accurate modeling over insight, but it seems to me that both notions have always coexisted as part of doing science. To test that, I consulted the epitome of doing science, namely Science. I looked at the current issue and chose a title and abstract at random: Chlorinated Indium Tin Oxide Electrodes with High Work Function for Organic Device Compatibility In organic light-emitting diodes (OLEDs), a stack of multiple organic layers facilitates charge flow from the low work function [~4.7 electron volts (eV)] of the transparent electrode (tin-doped indium oxide, ITO) to the deep energy levels (~6 eV) of the active light-emitting organic materials. We demonstrate a chlorinated ITO transparent electrode with a work function of >6.1 eV that provides a direct match to the energy levels of the active light-emitting materials in state-of-the art OLEDs. A highly simplified green OLED with a maximum external quantum efficiency (EQE) of 54% and power efficiency of 230 lumens per watt using outcoupling enhancement was demonstrated, as were EQE of 50% and power efficiency of 110 lumens per watt at 10,000 candelas per square meter. It certainly seems that this article is much more focused on "accurately modeling the world" than on "providing insight." The paper does indeed fit in to a body of theories, but it is mostly reporting on specific experiments and the results obtained from them (e.g. efficiency of 54%). I then looked at all the titles and abstracts from the current issue of Science: Comparative Functional Genomics of the Fission Yeasts Dimensionality Control of Electronic Phase Transitions in Nickel-Oxide Superlattices Competition of Superconducting Phenomena and Kondo Screening at the Nanoscale Chlorinated Indium Tin Oxide Electrodes with High Work Function for Organic Device
11-6-1 9:43
Compatibility Probing Asthenospheric Density, Temperature, and Elastic Moduli Below the Western United States Impact of Polar Ozone Depletion on Subtropical Precipitation Fossil Evidence on Origin of the Mammalian Brain Industrial Melanism in British Peppered Moths Has a Singular and Recent Mutational Origin The Selaginella Genome Identifies Genetic Changes Associated with the Evolution of Vascular Plants Chromatin "Prepattern" and Histone Modifiers in a Fate Choice for Liver and Pancreas Spatial Coupling of mTOR and Autophagy Augments Secretory Phenotypes Diet Drives Convergence in Gut Microbiome Functions Across Mammalian Phylogeny and Within Humans The Toll-Like Receptor 2 Pathway Establishes Colonization by a Commensal of the Human Microbiota A Packing Mechanism for Nucleosome Organization Reconstituted Across a Eukaryotic Genome Structures of the Bacterial Ribosome in Classical and Hybrid States of tRNA Binding and did the same for the current issue of Cell: Mapping the NPHP-JBTS-MKS Protein Network Reveals Ciliopathy Disease Genes and Pathways Double-Strand Break Repair-Independent Role for BRCA2 in Blocking Stalled Replication Fork Degradation by MRE11 Establishment and Maintenance of Alternative Chromatin States at a Multicopy Gene Locus An Epigenetic Signature for Monoallelic Olfactory Receptor Expression Distinct p53 Transcriptional Programs Dictate Acute DNA-Damage Responses and Tumor Suppression An ADIOL-ER-CtBP Transrepression Pathway Negatively Regulates Microglia-Mediated Inflammation A Hormone-Dependent Module Regulating Energy Balance Class IIa Histone Deacetylases Are Hormone-Activated Regulators of FOXO and Mammalian Glucose Homeostasis and for the 2010 Nobel Prizes in science: Physics: for groundbreaking experiments regarding the two-dimensional material graphene Chemistry: for palladium-catalyzed cross couplings in organic synthesis Physiology or Medicine: for the development of in vitro fertilization My conclusion is that 100% of these articles and awards are more about "accurately modeling the world" than they are about "providing insight," although they all have some theoretical insight component as well. I recognize that judging one way or the other is a difficult ill-defined task, and that you shouldn't accept my judgements, because I have an inherent bias. (I was considering running an experiment on Mechanical Turk to get an unbiased answer, but those familiar with Mechanical Turk told me these questions are probably too hard. So you the reader can do your own experiment and see if you agree.)
http://norvig.com/chomsky.html
7 29
11-6-1 9:43
What doesn't Chomsky like about statistical models?

I said that statistical models are sometimes confused with probabilistic models; let's first consider the extent to which Chomsky's objections are actually about probabilistic models. In 1969 he famously wrote: But it must be recognized that the notion of "probability of a sentence" is an entirely useless one, under any known interpretation of this term. His main argument being that, under any interpretation known to him, the probability of a novel sentence must be zero, and since novel sentences are in fact generated all the time, there is a contradiction. The resolution of this contradiction is of course that it is not necessary to assign a probability of zero to a novel sentence; in fact, with current probabilistic models it is well-known how to assign a non-zero probability to novel occurrences, so this criticism is invalid, but was very influential for decades. Previously, in Syntactic Structures (1957) Chomsky wrote: I think we are forced to conclude that ... probabilistic models give no particular insight into some of the basic problems of syntactic structure. In the footnote to this conclusion he considers the possibility of a useful probabilistic/statistical model, saying "I would certainly not care to argue that ... is unthinkable, but I know of no suggestion to this effect that does not have obvious flaws." The main "obvious flaw" is this: Consider 1. 2. 3. 4. I never, ever, ever, ever, ... fiddle around in any way with electrical equipment. She never, ever, ever, ever, ... fiddles around in any way with electrical equipment. * I never, ever, ever, ever, ... fiddles around in any way with electrical equipment. * She never, ever, ever, ever, ... fiddle around in any way with electrical equipment.
No matter how many repetitions of "ever" you insert, sentences 1 and 2 are grammatical and 3 and 4 are ungrammatical. A probabilistic Markov-chain model with n states can never make the necessary distinction (between 1 or 2 versus 3 or 4) when there are more than n copies of "ever." Therefore, a probabilistic Markov-chain model cannot handle all of English. This criticism is correct, but it is a criticism of Markov-chain modelsit has nothing to do with probabilistic models (or trained models) at all. Moreover, since 1957 we have seen many types of probabilistic language models beyond the Markov-chain word models. These particular examples can be distinguished with a finite-state model that is not a chain, but other examples require more sophisticated models. The best studied is probabilistic context-free grammar (PCFG), which operates over trees, categories of words, and individual lexical items,and has none of the restrictions of finitestate models. We find that PCFGs are state-of-the-art for parsing performance and are easier to learn from data than categorical context-free grammars. Other types of probabilistic models cover semantic and discourse structures. Every probabilistic model is a superset of a deterministic model (because the deterministic model could be seen as a probabilistic model where the probabilities are restricted to be 0 or 1), so any valid criticism of probabilistic models would have to be because they are too expressive, not because they are not expressive enough. In Syntactic Structures, Chomsky introduces a now-famous example that is another
11-6-1 9:43
criticism of finite-state probabilistic models: Neither (a) 'colorless green ideas sleep furiously' nor (b) 'furiously sleep ideas green colorless', nor any of their parts, has ever occurred in the past linguistic experience of an English speaker. But (a) is grammatical, while (b) is not. Chomsky appears to be correct that neither sentence appeared in the published literature before 1955. I'm not sure what he meant by "any of their parts," but certainly two-word parts had occurred, for example: "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). But regardless of what is meant by "part," a statistically-trained finite-state model can in fact distinguish between these two sentences. Pereira (2001) showed that such a model, augmented with word categories and trained by expectation maximization on newspaper text, computes that (a) is 200,000 times more probable than (b). To prove that this was not the result of Chomsky's sentence itself sneaking into newspaper text, I repeated the experiment, using a much cruder model with Laplacian smoothing and no categories, trained over the Google Book corpus from 1800 to 1954, and found that (a) is about 10,000 times more probable. If we had a probabilistic model over trees as well as word sequences, we could perhaps do an even better job of computing degree of grammaticality. Now you can say that you are more interested in a binary grammatical/ungrammatical distinction, but it is undeniable that a probabilistic system that not only makes the binary distinction but also associated a degree of grammaticality is strictly more informative than a binary system. Furthermore, the statistical models are capable of delivering the judgment that both sentences are extremely improbable, when compared to, say, "Effective green products sell well." Chomsky's theory, being categorical, cannot make this distinction; all it can distinguish is grammatical/ungrammatical. Another part of Chomsky's objection is "we cannot seriously propose that a child learns the values of 109 parameters in a childhood lasting only 108 seconds." (Note that modern models are much larger than the 109 parameters that were contemplated in the 1960s.) But of course nobody is proposing that these parameters are learned one-by-one; the right way to do learning is to set large swaths of nearzero parameters simultaneously with a smoothing or regularization procedure, and update the highprobability parameters continuously as observations comes in. And noone is suggesting that Markov models by themselves are a serious model of human language performance. But I (and others) suggest that probabilistic, trained models are a better model of human language performance than are categorical, untrained models. And yes, it seems clear that an adult speaker of English does know billions of facts (for example, that one says "big game" rather than "large game" when talking about an important football game). These facts must somehow be encoded in the brain. It seems clear that probabilistic models are better for judging the likelihood of a sentence, or its degree of sensibility. But even if you are not interested in these factors and are only interested in the grammaticality of sentences, it still seems that probabilistic models do a better job at describing the
11-6-1 9:43
linguistic facts. The mathematical theory of formal languages defines a language as a set of sentences. That is, every sentence is either grammatical or ungrammatical; there is no need for probability in this framework. But natural languages are not like that. A scientific theory of natural languages must account for the many phrases and sentences which leave a native speaker uncertain about their grammaticality (see Chris Manning's article and its discussion of the phrase "as least as"), and there are phrases which some speakers find perfectly grammatical, others perfectly ungrammatical, and still others will flip-flop from one occasion to the next. Finally, there are usages which are rare in a language, but cannot be dismissed if one is concerned with actual data. For example, the verb quake is listed as intransitive in dictionaries, meaning that (1) below is grammatical, and (2) is not, according to a categorical theory of grammar. 1. The earth quaked. 2. ? It quaked her bowels. But (2) actually appears as a sentence of English. This poses a dilemma for the categorical theory. When (2) is observed we must either arbitrarily dismiss it as an error that is outside the bounds of our model (without any theoretical grounds for doing so), or we must change the theory to allow (2), which often results in the acceptance of a flood of sentences that we would prefer to remain ungrammatical. As Edward Sapir said in 1921, "All grammars leak." But in a probabilistic model there is no difficulty; we can say that quake has a high probability of being used intransitively, and a low probability of transitive use (and we can, if we care, further describe those uses through subcategorization). Steve Abney points out that probabilistic models are better suited for modeling language change. He cites the example of a 15th century Englishman who goes to the pub every day and orders "Ale!" Under a categorical model, you could reasonably expect that one day he would be served eel, because the great vowel shift flipped a Boolean parameter in his mind a day before it flipped the parameter in the publican's. In a probabilistic framework, there will be multiple parameters, perhaps with continuous values, and it is easy to see how the shift can take place gradually over two centuries. Thus it seems that grammaticality is not a categorical, deterministic judgment but rather an inherently probabilistic one. This becomes clear to anyone who spends time making observations of a corpus of actual sentences, but can remain unknown to those who think that the object of study is their own set of intuitions about grammaticality. Both observation and intuition have been used in the history of science, so neither is "novel," but it is observation, not intuition that is the dominant model for science. Now let's consider what I think is Chomsky's main point of disagreement with statistical models: the tension between "accurate description" and "insight." This is an old distinction. Charles Darwin (biologist, 18091882) is best known for his insightful theories but he stressed the importance of accurate description, saying "False facts are highly injurious to the progress of science, for they often endure long; but false views, if supported by some evidence, do little harm, for every one takes a salutary pleasure in proving their falseness." More recently, Richard Feynman (physicist, 19181988) wrote "Physics can progress without the proofs, but we can't go on without the facts." On the other side, Ernest Rutherford (physicist, 18711937) disdained mere
10 29
11-6-1 9:43
On the other side, Ernest Rutherford (physicist, 18711937) disdained mere description, saying "All science is either physics or stamp collecting." Chomsky stands with him: "You can also collect butterflies and make many observations. If you like butterflies, that's fine; but such work must not be confounded with research, which is concerned to discover explanatory principles." Acknowledging both sides is Robert Millikan (physicist, 18681953) who said in his Nobel acceptance speech "Science walks forward on two feet, namely theory and experiment ... Sometimes it is one foot that is put forward first, sometimes the other, but continuous progress is only made by the use of both."
Butterflies
The two cultures

After all those distinguished scientists have weighed in, I think the most relevant contribution to the current discussion is the 2001 paper by Leo Breiman (statistician, 19282005), Statistical Modeling: The Two Cultures. In this paper Breiman, alluding to C.P. Snow, describes two cultures: First the data modeling culture (to which, Breiman estimates, 98% of statisticians subscribe) holds that nature can be described as a black box that has a relatively simple underlying model which maps from input variables to output variables (with perhaps some random noise thrown in). It is the job of the statistician to wisely choose an underlying model that reflects the reality of nature, and then use statistical data to estimate the parameters of the model.
Leo Breiman
Second the algorithmic modeling culture (subscribed to by 2% of statisticians and many researchers in biology, artificial intelligence, and other fields that deal with complex phenomena), which holds that nature's black box cannot necessarily be described by a simple model. Complex algorithmic approaches (such as support vector machines or boosted decision trees or deep belief networks) are used to estimate the function that maps from input to output variables, but we have no expectation that the form of the function that emerges from this complex algorithm reflects the true underlying nature. It seems that the algorithmic modeling culture is what Chomsky is objecting to most vigorously. It is not just that the models are statistical (or probabilistic), it is that they produce a form that, while accurately modeling reality, is not easily interpretable by humans, and makes no claim to correspond to the generative process used by nature. In other words, algorithmic modeling describes what does happen, but it doesn't answer the question of why. Breiman's article explains his objections to the first culture, data modeling. Basically, the conclusions made by data modeling are about the model, not about nature. (Aside: I remember in 2000 hearing James Martin, the leader of the Viking missions to Mars, saying that his job as a spacecraft engineer was not to land on Mars, but to land on the model of Mars provided by the geologists.) The problem is, if the model does not emulate nature well, then the conclusions may be wrong. For example, linear regression is one of the most powerful tools in the statistician's toolbox. Therefore, many analyses start out with "Assume the data are generated by a linear model..." and lack sufficient analysis of what happens if the data are not in fact generated that way. In addition, for complex problems there are usually many alternative good models, each with very similar measures of goodness of fit. How is the
11-6-1 9:43
data modeler to choose between them? Something has to give. Breiman is inviting us to give up on the idea that we can uniquely model the true underlying form of nature's function from inputs to outputs. Instead he asks us to be satisfied with a function that accounts for the observed data well, and generalizes to new, previously unseen data well, but may be expressed in a complex mathematical form that may bear no relation to the "true" function's form (if such a true function even exists). Chomsky takes the opposite approach: he prefers to keep a simple, elegant model, and give up on the idea that the model will represent the data well. Instead, he declares that what he calls performance datawhat people actually dois off limits to linguistics; what really matters is competencewhat he imagines that they should do. In January of 2011, television personality Bill O'Reilly weighed in on more than one culture war with his statement "tide goes in, tide goes out. Never a miscommunication. You can't explain that," which he proposed as an argument for the existence of God. O'Reilly was ridiculed by his detractors for not knowing that tides can be readily explained by a system of partial differential equations describing the gravitational interaction of sun, earth, and moon (a fact that was first Bill O'Reilly worked out by Laplace in 1776 and has been considerably refined since; when asked by Napoleon why the creator did not enter into his calculations, Laplace said "I had no need of that hypothesis."). (O'Reilly also seems not to know about Deimos and Phobos (two of my favorite moons in the entire solar system, along with Europa, Io, and Titan), nor that Mars and Venus orbit the sun, nor that the reason Venus has no moons is because it is so close to the sun that there is scant room for a stable lunar orbit.) But O'Reilly realizes that it doesn't matter what his detractors think of his astronomical ignorance, because his supporters think he has gotten exactly to the key issue: why? He doesn't care how the tides work, tell him why they work. Why is the moon at the right distance to provide a gentle tide, and Laplace exert a stabilizing effect on earth's axis of rotation, thus protecting life here? Why does gravity work the way it does? Why does anything at all exist rather than not exist? O'Reilly is correct that these questions can only be addressed by mythmaking, religion or philosophy, not by science. Chomsky has a philosophy based on the idea that we should focus on the deep whys and that mere explanations of reality don't matter. In this, Chomsky is in complete agreement with O'Reilly. (I recognize that the previous sentence would have an extremely low probability in a probabilistic model trained on a newspaper or TV corpus.) Chomsky believes a theory of language should be simple and understandable, like a linear regression model where we know the underlying process is a straight line, and all we have to do is estimate the slope and intercept. For example, consider the notion of a pro-drop language from Chomsky's Lectures on Government and Binding (1981). In English we say, for example, "I'm hungry," expressing the pronoun "I". But in Spanish, one expresses the same thought with "Tengo hambre" (literally "have hunger"), dropping the pronoun "Yo". Chomsky's theory is that there is a "pro-drop parameter" which is "true" in Spanish and "false" in English, and that once we discover the small set of parameters that describe all languages, and the values of those parameters for each language, we will have achieved true understanding. The problem is that reality is messier than this theory. Here are some dropped pronouns in English:
11-6-1 9:43
"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)
Dana Carvey
Linguists can argue over the interpretation of these facts for hours on end, but the diversity of language seems to be much more complex than a single Boolean value for a pro-drop parameter. We shouldn't accept a theoretical framework that places a priority on making the model simple over making it accurately reflect reality. From the beginning, Chomsky has focused on the generative side of language. From this side, it is reasonable to tell a non-probabilistic story: I know definitively the idea I want to expressI'm starting from a single semantic formthus all I have to do is choose the words to say it; why can't that be a deterministic, categorical process? If Chomsky had focused on the other side, interpretation, as Claude Shannon did, he may have changed his tune. In interpretation (such as speech recognition) the listener receives a noisy, ambiguous signal and needs to decide which of many possible intended messages is most likely. Thus, it is obvious that this is inherently a probabilistic problem, as was recognized early on by all researchers in speech recognition, and by scientists in other fields that do interpretation: the astronomer Laplace said in 1819 "Probability theory is nothing more than common sense reduced to calculation," and the physicist James Maxwell said in 1850 "The true logic for this world is the calculus of Probabilities, which takes account of the magnitude of the probability which is, or ought to be, in a reasonable man's mind." Finally, one more reason why Chomsky dislikes statistical models is that they tend to make linguistics an empirical science (a science about how people actually use language) rather than a mathematical science (an investigation of the mathematical properties of models of formal language). Chomsky prefers the later, as evidenced by his statement in Aspects of the Theory of Syntax (1965): Linguistic theory is mentalistic, since it is concerned with discovering a mental reality underlying actual behavior. Observed use of language ... may provide evidence ... but surely cannot constitute the subject-matter of linguistics, if this is to be a serious discipline. I can't imagine Laplace saying that observations of the planets cannot constitute the subject-matter of orbital mechanics, or Maxwell saying that observations of electrical charge cannot constitute the subject-matter of electromagnetism. It is true that physics considers idealizations that are abstractions from the messy real world. For example, a class of mechanics problems ignores friction. But that doesn't mean that friction is not considered part of the subject-matter of physics. So how could Chomsky say that observations of language cannot be the subjectmatter of linguistics? It seems to come from his viewpoint as a Platonist and a Rationalist and perhaps a bit of a Mystic. As in Plato's analogy of the cave, Chomsky thinks we should focus on the ideal, abstract forms that underlie language, Plato's cave not on the superficial manifestations of language that happen to be perceivable in the real world. That is why he is not interested in language performance. But Chomsky, like Plato, has
11-6-1 9:43
to answer where these ideal forms come from. Chomsky (1991) shows that he is happy with a Mystical answer, although he shifts vocabulary from "soul" to "biological endowment." Plato's answer was that the knowledge is 'remembered' from an earlier existence. The answer calls for a mechanism: perhaps the immortal soul ... rephrasing Plato's answer in terms more congenial to us today, we will say that the basic properties of cognitive systems are innate to the mind, part of human biological endowment. It was reasonable for Plato to think that the ideal of, say, a horse, was more important than any individual horse we can perceive in the world. In 400BC, species were thought to be eternal and unchanging. We now know that is not true; that the horses on another cave wallin Lascauxare now extinct, and that current horses continue to evolve slowly over time. Thus there is no such thing as a single ideal eternal "horse" form.
Lascaux Horse
We also now know that language is like that as well: languages are complex, random, contingent biological processes that are subject to the whims of evolution and cultural change. What constitutes a language is not an eternal ideal form, represented by the settings of a small number of parameters, but rather is the contingent outcome of complex processes. Since they are contingent, it seems they can only be analyzed with probabilistic models. Since people have to continually understand the uncertain. ambiguous, noisy speech of others, it seems they must be using something like probabilistic reasoning. Chomsky for some reason wants to avoid this, and therefore he must declare the actual facts of language use out of bounds and declare that true linguistics only exists in the mathematical realm, where he can impose the formalism he wants. Then, to get language from this abstract, eternal, mathematical realm into the heads of people, he must fabricate a mystical facility that is exactly tuned to the eternal realm. This may be very interesting from a mathematical point of view, but it misses the point about what language is, and how it works.
Thanks
Thanks to Ann Farmer, Fernando Pereira, Dan Jurafsky, Hal Varian, and others for comments and suggestions on this essay.
Annotated Bibliography
1. Abney, Steve (1996) Statistical Methods and Linguistics, in Klavans and Resnik (eds.) The Balancing Act: Combining Symbolic and Statistical Approaches to Language, MIT Press. An excellent overall introduction to the statistical approach to language processing, and covers some ground that is not addressed often, such as language change and individual differences. 2. Breiman, Leo (2001) Statistical Modeling: The Two Cultures, Statistical Science, Vol. 16, No. 3, 199-231. Breiman does a great job of describing the two approaches, explaining the benefits
11-6-1 9:43
of his approach, and defending his points in the vary interesting commentary with eminent statisticians: Cox, Efron, Hoadley, and Parzen. 3. Chomsky, Noam (1956) Three Models for the Description of Language, IRE Transactions on Information theory (2), pp. 113-124. Compares finite state, phrase structure, and transformational grammars. Introduces "colorless green ideas sleep furiously." 4. Chomsky, Noam (1967) Syntactic Structures, Mouton. A book-length exposition of Chomsky's theory that was the leading exposition of linguistics for a decade. Claims that probabilistic models give no insight into syntax. 5. Chomsky, Noam (1969) Some Empirical Assumptions in Modern Philosophy of Language, in Philosophy, Science and Method: Essays in Honor or Ernest Nagel, St. Martin's Press. Claims that the notion "probability of a sentence" is an entirely useless notion. 6. Chomsky, Noam (1981) Lectures on government and binding, de Gruyer. A revision of Chomsky's theory; this version introduces Universal Grammar. We cite it for the coverage of parameters such as pro-drop. 7. Chomsky, Noam (1991) Linguistics and adjacent fields: a personal view, in Kasher (ed.), A Chomskyan Turn, Oxford. I found the Plato quotes in this article, published by the Communist Party of Great Britain, and apparently published by someone with no linguistics training whatsoever, but with a political agenda. 8. Gold, E. M. (1967) Language Identification in the Limit, Information and Control, Vol. 10, No. 5, pp. 447-474. Gold proved a result in formal language theory that we can state (with some artistic license) as this: imagine a game between two players, guesser and chooser. Chooser says to guesser, "Here is an infinite number of languages. I'm going to choose one of them, and start reading sentences to you that come from that language. On your Nth birthday there will be a True-False quiz where I give you 100 sentences you haven't heard yet, and you have to say whether they come from the language or not." There are some limits on what the infinite set looks like and on how the chooser can pick sentences (he can be deliberately tricky, but he can't just repeat the same sentence over and over, for example). Gold's result is that if the infinite set of languages are all generated by context-free grammars then there is no strategy for guesser that guarantees she gets 100% correct every time, no matter what N you choose for the birthday. This result was taken by Chomsky and others to mean that it is impossible for children to learn human languages without having an innate "language organ." As Johnson (2004) and others show, this was an invalid
11-6-1 9:43
conclusion; the task of getting 100% on the quiz (which Gold called language identification) really has nothing in common with the task of language acquisition performed by children, so Gold's Theorem has no relevance. 9. Horning, J. J. (1969) A study of grammatical inference, Ph.D. thesis, Stanford Univ. Where Gold found a negative resultthat context-free languages were not identifiable from examples, Horning found a positive resultthat probabilistic context-free languages are identifiable (to within an arbitrarily small level of error). Nobody doubts that humans have unique innate capabilities for understanding language (although it is unknown to what extent these capabilities are specific to language and to what extent they are general cognitive abilities related to sequencing and forming abstractions). But Horning proved in 1969 that Gold cannot be used as a convincing argument for an innate language organ that specifies all of language except for the setting of a few parameters. 10. Johnson, Kent (2004) Gold's Theorem and cognitive science, Philosophy of Science, Vol. 71, pp. 571-592. The best article I've seen on what Gold's Theorem actually says and what has been claimed about it (correctly and incorrectly). Concludes that Gold has something to say about formal languages, but nothing about child language acquisition. 11. Lappin, Shalom and Shieber, Stuart M. (2007) Machine learning theory and practice as a source of insight into universal grammar., Journal of Linguistics, Vol. 43, No. 2, pp. 393-427. An excellent article discussing the poverty of the stimulus, the fact that all models have bias, the difference between supervised and unsupervised learning, and modern (PAC or VC) learning theory. It provides alternatives to the model of Universal Grammar consisting of a fixed set of binary parameters. 12. Manning, Christopher (2002) Probabilistic Syntax, in Bod, Hay, and Jannedy (eds.), Probabilistic Linguistics, MIT Press. A compelling introduction to probabilistic syntax, and how it is a better model for linguistic facts than categorical syntax. Covers "the joys and perils of corpus linguistics." 13. Norvig, Peter (2007) How to Write a Spelling Corrector, unpublished web page. Shows working code to implement a probabilistic, statistical spelling correction algorithm. 14. Norvig, Peter (2009) Natural Language Corpus Data, in Seagran and Hammerbacher (eds.), Beautiful Data, O'Reilly. Expands on the essay above; shows how to implement three tasks: text segmentation, cryptographic decoding, and spelling correction (in a slightly more complete form
11-6-1 9:43
than the previous essay). 15. Pereira, Fernando (2002) Formal grammar and information theory: together again?, in Nevin and Johnson (eds.), The Legacy of Zellig Harris, Benjamins. When I set out to write the page you are reading now, I was concentrating on the events that took place in Cambridge, Mass., 4800 km from home. After doing some research I was surprised to learn that the authors of two of the three best articles on this subject sit within a total of 10 meters from my desk: Fernando Pereira and Chris Manning. (The third, Steve Abney, sits 3700 km away.) But perhaps I shouldn't have been surprised. I remember giving a talk at ACL on the corpus-based language models used at Google, and having Fernando, then a professor at U. Penn., comment "I feel like I'm a particle physicist and you've got the only super-collider." A few years later he moved to Google. Fernando is also famous for his quote "The older I get, the further down the Chomsky Hierarchy I go." His article here covers some of the same ground as mine, but he goes farther in explaining the range of probabilistic models available and how they are useful. 16. Plato (c. 380BC) The Republic. Cited here for the analogy of the cave. 17. Shannon, C.E. (1948) A Mathematical Theory of Communication, The Bell System Technical Journal, Vol. 27, pp. 379-423. An enormously influential article that started the field of information theory and introduced the term "bit" and the noisy channel model, demonstrated successive ngram approximations of English, described Markov models of language, defined entropy with respect to these models, and enabled the growth of the telecommunications industry. Peter Norvig
and 6 others liked this.
Add New Comment

Optional: Login below.
17 29
11-6-1 9:43
Image
Post as
Showing 21 comments
Sort by
Popular now
Subscribe by email
Subscribe by RSS
guest 9 hours ago

Found a small typo: "concise tree-structure model without probabilities cannot model of all language" did you mean model all of language I'll be looking for my 2.56 in the mail.
Like Reply
Barbara H Partee 22 hours ago

I'd like to suggest a distinction that might be useful to make in addressing the formal linguists. You wrote, "But the vast majority of people who study interpretation tasks, such as speech recognition, quickly see that interpretation is an inherently probabilistic problem: given a stream of noisy input to my ears, what did the speaker most likely mean?" That's one kind of interpretation task (I'll call it the 'real-world interpretation task'), and I guess it's the central one for applications. Another interpretation problem is this: given a stream of input, what are the possible things the speaker could have meant? (Of course it's non-trivial to formulate this question in a useful way -- it's natural to make idealizations both about the input stream and about the 'competence' of the speaker, but I don't want to get into such topics, since that could seem to be suggesting that this is about whether to ignore real-world data or not.) I'd like to see inclusion of a version of the interpretation problem that reflects my own work as a working formal semanticist and is not inherently more probabilistic than the formal 'generation task' (which, by the way, has very little in common with the real-world sentence production task, a task that is probably just as probabilistic as the real-world interpretation task). You've said in your replies to earlier comments that you plan to write a follow-up essay centered on the thesis that the best model for explaining language is a probabilistic model. I think that's the best strategy. And in the realm of interpretation, we (formal linguists) want to explain (among other things) the possible scopes of quantifiers, the possible interpretations for anaphoric expressions in various configurations, etc. Some of our favorite puzzles concern negative data -- why certain sentences cannot have certain meanings that would seem to be quite plausible. It's the possibility that probabilistic methods could be helpful for the problems I work on as well as for the problems you work on that makes me prick up my ears. Of course you may want to try to convince me that I should be working on different problems if I want to explain natural language, but I feel like our community has made real progress, with real discoveries, and even a few robust results, so I'd be much happier at the prospect of merging efforts than at the prospect of having to make an either-or choice. p.s. Was I correct when I said in that same round table session that I doubted that Google makes any use of any formal semantic analyses in its responses to search queries? (That was my way of acknowledging that at least what I do myself seems very far away from direct usability in any real-world applications. But I know there are some computational applications of formal semantics, I just don't know much about them.)
Like Reply
18 29
11-6-1 9:43
Peter Norvig
18 hours ago in reply to Barbara H Partee
Thanks, Barbara. I meant to include some of your very interesting comments at the Symposium, but couldn't find room -- maybe for the next piece, as you suggest. I support your notion of merging efforts. I think different communities will have different priorities, and thus different emphasis on what they do, but there is a chance to bring them together where there is overlap.Google does not use formal semantic analysis of the kind you work on for search queries, because the language we see in search queries is so far removed from full grammatical sentences. (If you see mostly 3 words or so for a query, you don't have to worry about nested embedded clauses, etc.) We are starting to experiment with more serious semantic analysis for understanding text in documents which does have grammatical structure.
Barbara H Partee liked this Like Reply
benjamin 1 day ago

A very interesting article, here are some questions I have and I would be very thankful if you found the time to answer them: "Third, I believe that language understanding is an inherently a process of reasoning with uncertainty, and that since probability theory is axiomatically the way to deal with uncertainty, I find it inescapable that probability should be used to model language." Would you mind elaborating on what exactly you mean by this? In particular, what is the connection between the fact that 'understanding is an inherently a [sic] process of reasoning with uncertainty' and your assertion that 'probability should be used to model language'. Whatever language is, and whatever the process of understanding is, they seem to be to different kinds of things. So I would agree that we should use probability to model language understanding (and, say, language learning) but I do not quite see why this should argue for using probability to model language. By the way. what IS a language? "languages are complex, random, contingent biological processes that are subject to the whims of evolution and cultural change." ? Quite frankly, I don't find that very enlightening, and in particular I do not understand what 'random' is supposed to mean here. (I have no trouble understanding 'random' in contexts such as 'random mutation') A related question is: what kind of probability are you talking about? Perhaps this seems like an outlandish question, but if one wants to actually do Science instead of 'simply' use probabilistic models for engineering purposes (not that anything is trivial or wrong about that), one ought to be clear how notions such as P('colorless green ideas sleep furiously') are to be understood. Otherwise we don't know which parts of our objective reality are supposed to make our 'scientific' assertions such as 'P(sentence/meaning/what-have-you) is x' true. What do you take to be the truthmakers for such sentences? Facts about the linguistic behavior of some community or speaker (say, how often they actually use a sentence / would have used a sentence / will use a sentence / ...)? "Observed use of language ... may provide evidence ... but surely cannot constitute the subject-matter of linguistics, if this is to be a serious discipline. " [Chomsky 1965] Here is another take on how to understand this quote (admittedly, Chomsky does not seem to have pressed this point recently, but it seems reasonable enough to me to be worth considering, and it seems to me to be exactly the point he made in his 1959 review of 'Verbal Behavior'), related to my previous question: If observed use of language (instead of the 'competence' underlying this behavior) was the subject matter of linguistics, what kinds of generalizations would linguistics aim for? Generalizations of 'language use', it would seem to
19 29
11-6-1 9:43
me, and that would seem to me to finally include generalizations about the causation of specific instances of linguistic behavior. (most probably probabilistic generalizations?) If that were so, Chomsky 1959 (http://www.chomsky.info/articl... has a nice discussion of general problems a science which aims at explaining actual behavior has to deal with.
(One minor thing - the quote "Linguistic Theory is mentalistic..." is not from Syntactic Structures, but from Aspects of the Theory of Syntax (the link you set is correct, as is the year you give, 1965, but you don't list it in the bibliography).)
Like
Reply
timmenzies 2 days ago

This article, and the comments below, are *perfect* fuel for an intro lecture to data mining subject. Thanks to Norvig, and all the commenters, for their work. But if someone offers a dichotomy in an argument, consider that such divisions may be too rigid and look for some middle ground. To that end, I offer the following. Note that I am not talking about linguistics in particular but rather I am commenting on the value of data mining in general. Consider Wikipedia's list of human cognitive biases (http://en.wikipedia.org/wiki/L... A *huge* list of how people can get it wrong wrong wrong when they analyze data. Now consider another Norvig essay. http://norvig.com/experiment-d... Look for " guess which of these four squares". My take from that example is that, if you look for structures in a random sample, you will find it (albiet at some low prob). From these, I offer two lessons: - Lesson #1: humans can be kinda dumb and we can't rely on just their insights. - Lesson #2: dumb algorithms can be kinda dumb so we can't rely on just their insights. So, how do we know that when we try to generalize from the past that "all is not a woven web of guesses" (http://goo.gl/wfAx2)? Well, how about hybrid methods? Communities of agents (some that use proteins, some that use silicon), each with their own search biases and tendencies to make their particular kinds mistakes. Then we get them together in committees and let each shed light on the confusion of the others. So not "Everything is basic theory" or "everything is probabilistic inference" but some middle ground where ensembles of data miners & people share a similar ontology. Where insights & conclusions from one group can be read by the other (and by "read" I mean, audited, maybe refuted or improved). For the geeks: bring on the ensembles boosting and multi-dimensional visualizations and feature/instance selection to reduce the space that people have to discuss. Yes, have probabilistic algorithms since they scale but use those to rank short categorical rules that humans can understand. Favor minimal contrast set learners over classifiers since the difference between things can be shorter to read than a detailed description of the things themselves. E.g. http://goo.gl/a4BfR.
20 29
11-6-1 9:43
Like
Reply
Peter Norvig
Good comment, Tim.
2 days ago in reply to timmenzies
The word "essay" means "to try" and part of what I'm doing here is try to make sense of a complex history of ideas, and part is trying to understand and overcome my own cognitive biases -- I know I have them, but I have to try very hard to recognize and reconcile them. I'm about to take off for vacation with no Internet access, so expect me to go silent on replies.And next time I'm in the WVU area, maybe I can participate in a class discussion in your class.-Peter
Like Reply
timmenzies 2 days ago in reply to Peter Norvig

Guest talk at WVU? W'ed be honored.
Like Reply
Jonathan Miller 2 days ago

But surely the question is what did he actually say? I don't understand how you can found an entire paper on what you 'take to mean.'
Like
Reply
kathy 2 days ago

Whatever the merits of the dichotomy drawn (and I don't think we really know enough about the workings of the mind to decide that now) this is a wonderful discussion; thanks to all participants for their time on it. Now -- about the tides :) While can write equations for the tides and even the weather, producing solutions is often very hard. In the case of the tides, you have to deal with all kinds of irregularities, from the tilt of the earth, the depth of the seabed, channels, coasts .... . A central theory and perturbations from it. In the end, you understand a certain kind of 'why' but to produce accurate predictions, you'll rely on fitting data. To say that science is either one or the other is, I think, overstating the case.
Like Reply
Vijay L 2 days ago

> if Vijay thinks I was condescending, I sincerely apologize. I didn't.
11-6-1 9:43
Like
Reply
majorclanger 2 days ago

Fun essay on language, statistics and the nature of science
Like Reply
Vijay L 2 days ago

Chomsky has always called on linguistics to be approached as an empirical science. He has discussed this on several occasions with, for instance, Quine. See New Horizons in the Study of Language and Mind, for one. The problem with all stochastic and probabilistic models is that they achieve a similar output as a human being but they are hardly accurate and definitely not reflective of what we do naturally. An example would be Deep Blue and other machines. They achieve the exact same output as humans -- good chess gameplay -- but they give absolutely no insight into how a human being actually plays chess. Even if all stochastic/probabilistic models were to achieve 100% accuracy in, say, translating texts, if they don't provide any insight on human understanding, "commit it then to the flames" as Hume put it.
Like Reply
Peter Norvig
2 days ago in reply to Vijay L
I'm sorry you didn't understand my article, Vijay; I'll try to improve it. I understand that Chomsky has called on Linguistics to be empirical at times, but when he says "Observed use of language ... cannot constitute the subject-matter of linguistics" he is being about as anti-empirical as one could possibly be. You make an amazing claim about "all stochastic and probabilistic models." Have you examined all such models? If your claim is true, then no logical (or categorical) model can be reflective of what we do, either, because every logical (or categorical) model is a special case of a probabilistic model (the case where all the probabilities are constrained to be 0 or 1. I think your mistake is that you see some probabilistic models that do not attempt to model the underlying processes, and you think every model must be the same. That is false. I would say that because people deal with uncertainty, and do so in a somewhat rational way some of the time, only a probabilistic model can reflect what they are doing, because only a probabilistic model (or a model that acts the same as a probabilistic model) can be rational in the face of uncertainty.
Vijay L and 2 more liked this Like Reply
Stephen Conn 2 days ago in reply to Peter Norvig

You are astonishingly condescending for someone speaking from a position of such profound ignorance. The commenter did not say that he did not understand your article. He challenged your claim that Chomsky is against empiricism by providing direct citations to works in which Chomsky discusses the empirical nature of linguistics. I doubt that you are really unable to understand the quote you've now twice attributed to Chomsky, but I'll elaborate on it since you insist on making it appear as
11-6-1 9:43
though you are. Chomsky is saying in the quote that the task of a theory of language is to explain linguistic competence in light of noisy linguistic data. You seem to have no trouble understanding that language is "noisy" when you cite Shannon; yet when Chomsky makes what is essentially the exact same claim, you deride him as a mystic and compare him to Plato as though it's an insult. To say that linguistic data are noisy in this context means that as a scientist it is your task to decide which portions of a set of data are worth explanation and further investigation. This claim has precedent in virtually every scientific investigation to take place in the history of science. To cite just one famous example, you could consider reading about Galileo's study of projectile motion in Whewell 1837. Galileo argued that the natural motion of a projectile was parabolic, despite the fact that no such motion had ever been observed. Scientists before him had analyzed the observed motion and had modeled it closely. Their models, in a stroke of ontological nihilism that is a propos, failed to get at the distinction between the aspects of the motion caused by gravity and those caused by air resistance. Galileo rejected them precisely because they were too close to the empirical data. The history of science resoundingly supports Chomsky's claim, which is that accurate simulation of some complex natural phenomenon does not in itself constitute (successful) scientific thought. The heart of science is indeed explanation, not modeling, as you bizarrely suggest. The words physics and nature themselves are from the Greek physis and Latin natura, which refer to the mysterious push or agency that drives a plant to grow up out of the ground. The beginnings of modern scientific thought are characterized by qualities which are precisely the opposite of those which you claim are important to science. We see in Ancient Egypt and Babylon intricately detailed celestial models; Stonehenge was a wildly precise astronomical instrument in its time. But it is the early Greeks who mark a break in the traditions of human thought and language, and whom we therefore credit as the progenitors of science, because they sought to explain the behavior of the natural world, and created oral and written traditions in support of this goal. Stonehenge was a ceremonial toy. Its creators were not great scientists. (See Casper and Noer 1972 for a good discussion of this distinction.) It is poetic for our discussion that Thales is often pointed to in particular as being the father of modern science, given how poorly the only "model" attributed him "fit the data." The notion that everything is derived from water is actually quite sweet--one can imagine pointing to the silt at the mouth of a river as evidence that land emerges from water, or the bubbles that emerge as it boils as evidence that air is derived from it. It was an earnest attempt to explain the natural world by a man much smarter than you or me, and it was significant not because it was a good model, but because it marked a change in the way that people talked about the natural world--the beginning of the search for better and better explanations for its phenomena, its agency, its "push." (See Rosenblatt 2010.) All Chomsky is saying is that he feels present statistical language models do not get at this push. This is a reasonable claim made in earnest. Your comparison to Bill O'Reilly is again bizarre and clearly disingenuous. Shall we deride Galileo as well for going after the "deep why" that explains the non-parabolic motion of a projectile?
Peter Norvig and 6 more liked this Like Reply
Peter Norvig
2 days ago in reply to Stephen Conn
Stephen, thanks for the great reply -- impressively rich with historical background. First of all, if Vijay thinks I was condescending, I sincerely apologize. That wasn't my intent at all. I was attempting to blame myself for not being clear in my essay (not about Chomsky as empiricist, which is a matter of opinion, and Vijay is certainly welcome to his opinion as I am to mine) but about Vijay's statement about "all probabilistic models". I tried to show that Vijay has a factual/definitional misunderstanding of what probabilistic models are (again, admitting that it was my fault if I didn't explain it, not Vijay's fault): I tried to point out that logical models are a proper subset of probabilistic models, whereas Vijay seemed to think that they were disjoint sets.
23 29
11-6-1 9:43
Look, lots of very smart people are saying this essay is exactly right; that they've always felt that way, and it puts things into words perfectly. Other equally smart people think I'm a lunatic who doesn't know the first thing about language, science, or how to put together a coherent paragraph. How could there be such a discrepancy? I'll take for granted that all these people are reasoning rationally. So if they have reached different conclusions, it must be either because they know something different, or they have different goals. I'm still trying to figure out where the differences in knowledge and goals are, and you're definitely helping me understand a little, so I certainly appreciate that. Now for your arguments. I think there is a lot we agree on, and a lot where there is a difference in emphasis. You give a very nice history of Galileo. I agree with you that a scientist has to decide what to look at and what to put aside for later (or leave for someone else). The reason we all applaud Galileo is not just that he boldly decided to split motion into the component from gravity and the component from air resistance, but more so because he chose correctly. If he had said that the true motion was a straight line, and air resistance messed up that straight line, we would not be reveling in his history now. So I think our difference is mostly that you think Chomsky is like Galileo: he has focused on the interesting part (competence) and is close to getting it right. I have three hesitations: first, while I agree that understanding competence is very important, I also think performance is very important. Second, I don't think Chomsky is currently very close to getting competence right. (If he did have it almost right, I would be celebrating the accomplishment and would have no complaints whatsoever. Sure, I would want somebody to look at performance, but I wouldn't blame Chomsky for leaving it for someone else.) Third, I believe that language understanding is an inherently a process of reasoning with uncertainty, and that since probability theory is axiomatically the way to deal with uncertainty, I find it inescapable that probability should be used to model language. Next, about the word "science". Why is it that my viewpoint seems so clear-cut to me? The dictionary definition says "systematic study of the structure and behavior of the physical and natural world through observation and experiment." My claim was that science is a combination of gathering facts and making theories. Out of all the Science and Cell titles/abstracts I posted, not one commenter has said that they believe a single one of the articles is more about "theory" or "insight" than they are about "observation" and "accurate facts/modeling." I could also point to the 2010 Nobel Prizes in Physics (for groundbreaking experiments regarding the two-dimensional material graphene) and Chemistry (for palladium-catalyzed cross couplings in organic synthesis) and Physiology or Medicine (for the development of in vitro fertilization). Those all sound like they're strongly on the "observation/experiment/facts" (and "engineering") side; in fact I was surprised; I was expecting it to be more evenly balanced with the "theory/insight" side. (I'd be interested Stephen, what your take on those articles and Nobel is: are there any that are predominantly about theory/insight?) And yet you (along with many of my other critics) are feeling equally good about yourself; the answer is equally clearcut to you: "The history of science resoundingly supports Chomsky's claim." The best resolution of this discrepancy that I can come up with is that I'm focusing on the day-to-day activities of a scientist, and you're focusing on the longterm end goal. 100 years from now, nobody cares what you did in the lab last Thursday. They care about the big theories. Your paragraph about "the push" and your final paragraph helped me better see what is going on. My opinion is that Chomsky is headed down the wrong path and has chosen to ignore the tool of probabilistic reasoning that could set things in a better direction (both for achieving accuracy and insight). You think he is headed on the right path and has wisely avoided the distraction of an irrelevant tool. Like two fans of opposing sports fans who each think a close play went in favor of their team, we interpret things to fit our view. Anyway, you and the other critics have helped me realize that what I should write next should take the following tack: forget about performance for now, and forget about accuracy of a model. My thesis is that the best model for explaining language is a probabilistic model.
3 people liked this.
Like
Reply
24 29
11-6-1 9:43
Stephen Conn 2 days ago in reply to Peter Norvig

You are welcome--and thank you. I apologize for jumping the gun on Vijay's behalf. You are right that we agree on a lot. I think your three points in paragraph three are spot on. Linguistic performance is fascinating and I agree that it is worth studying scientifically; the state of theoretical linguistics as an empirical science is poor at best; and I don't disagree that probabilistic models are tools that we should use to understand and model language. I disagree with your conflation of Chomsky with an entire subfield of linguistics as though he is responsible for its state. And I would mention that there are many intellectual descendants and friends of Chomsky who do study performance, and do so fruitfully. Take a stroll around the Stata Center and you will trip over dozens of brilliant people who fit just that description. A former mentor of mine, Martin Hackl, is just one small example: http://shass.mit.edu/research/... These people are sharp as a tack and they make theoretical decisions on principled, well-reasoned grounds, not out of empirical blindness. I'm somewhat concerned with your certainty regarding the claim that any criticism of a probabilistic model is necessarily a criticism of a logical model because the latter are a proper subset of the former. Your argument has no legs, and you actually acknowledge that in your original post when you write that probabilistic models are indeed susceptible to criticism for being too expressive. By the reasoning in your reply to Vijay, must logical models not also be susceptible to any such criticism? You recognize correctly in your original post that the answer is no, they need not: there are features that we can point to that logical models possess that probabilistic models do not, and these features are without a doubt fair game in a debate about the utility of a model (If there were no such features, how would we distinguish the logical models as a (proper) subset?). You write in your post that probabilistic models could be validly criticized for being over-expressive, but when Vijay makes just this claim, you tell him that no such claim can be made without also applying it to logical models. That's silly. Moreover, meta-level criticism of probabilistic or logical models does not define membership in either set, and thus the subset relationship is incidental to our discussion. For example, I could make the assertion, "Peter Norvig defends the validity of probabilistic models for modeling language use in his latest blog post." Surely you would not also obligate me to the claim, "Peter Norvig defends the validity of logical models for modeling language use in his latest blog post," due to the above referenced subset relation? For set-theoretic purposes, praise and criticism are interchangeable: you can praise the superset without praising the subset just as Vijay or I can criticize the superset without criticizing the subset. Your discussion of Galileo's work, though, reveals more profoundly where we differ. Again your characterization of his work is almost poetic in its wrongheadedness and the way that it captures your misunderstanding of the work of science. It is perfect fodder for our debate. I follow Casper and Noer closely again. Galileo's work on projectile motion represented a break from the prevailing Aristotelian view that a projectile's natural motion is towards the center of the earth--i.e., the center of the Aristotelian universe--and that movement in any other direction is constrained movement that must be sustained with a force. Aristotle argued that natural motion "comes for free"--drop a stone and it naturally returns to the center of the earth--and that constrained motion requires some sustained cause in order to continue--push your stone and you had better keep pushing or it eventually rolls to a stop. Galileo obliterated the ancient distinction between natural and constrained motion. As an aside before we go on, pause for a moment to meditate on what I've just said. Aristotle lived two thousand years before Galileo. In life but most especially in science it is easy and tempting for us to dismiss those around us and those who came before us as just not having gotten it. This is tragic intellectual cowardice. There were many brilliant scientists between Aristotle and Galileo, but the explanatory power of Aristotle's physics was so overwhelming that it took two thousand years for the model to be altered by Galileo. Aristotle's model sounds
25 29
11-6-1 9:43
coarse and childish to us now only because we are lulled by assurances from authority into not examining our own lack of understanding, not because we are smarter people living in a smarter time. If anything, the opposite is true. But back to our discussion of the substance of his work. Galileo argued that, contrary to Aristotle and every day experience, motion imparted to an object is not spent, that but for the presence of friction, the cart given a single pull by its horse would continue to roll on and on forever. Galileo invented circular inertial motion. And it is from this upending of the model that Galileo made the argument that the natural motion of a projectile should be parabolic. The component of the motion parallel to the surface of the earth was not being spent or used up except for the presence of the air; the projectile experiences no horizontal acceleration, and its motion can be understood given its initial horizontal velocity and its uniform acceleration towards the center of the earth. Was Galileo right, and do we remember him in part for getting it right? I suppose so, but again you've missed the forest for the trees. Science is about arguments, not answers. We remember Galileo because he invented inertia as a model for explaining the natural motion of objects near the surface of the earth. He did get it right in a meaningful sense, but he told only a tiny part of the story, and even the part that he told would go on to be understood much more richly than he understood it--see Newton and Leibniz. I envision in your characterization of the event Galileo sitting down in front of a series of plots of data and declaring the parabola to have the best fit--the linear model was tempting, but it just wouldn't work! This is quaint and, as you acknowledge in your original post, unscientific. It misses the point so fundamentally. I don't have access to the latest issues of Science or Cell, but your claim about Nobel Laureates is strange and I have trouble believing you actually make it sincerely. Take a look at http://en.wikipedia.org/wiki/L... Even with my sliver of knowledge on the history of physics, I can quickly point to Lorentz, the Curies, Thomson, Michelson, van der Waals, Planck, Einstein, Bohr, Millikan, de Broglie, Heisenberg, Schrodinger, Dirac from the first quarter of the page as having undeniably made significant theoretical contributions to physics. And I'd be willing to bet that the names that I have left out are due to my ignorance rather than the atheoretical nature of their work. As to Science and Cell: of course observation and experimentation are essential facets of science. Your suggestion that any one claims otherwise is a straw man. Indeed it was largely careful experimentation that inspired Galileo to make the above contribution to physics--he was one of the first physicists to note via careful experiment that a projectile imparted a vertical velocity experiences a constant downward acceleration of g. But this does not mean that empirical accuracy is the only arbiter of scientific success. In fact, as Chomsky claims, this has historically been a relatively small measure of scientific success. There is a very similar discussion to the above in Rosenblatt 2004 (Science & Education 13: 379-388), but it pushes the ideas even further, and I think you might enjoy it. I have a digital copy if you'd like it.
2 people liked this. Like Reply
Vijay L 2 days ago in reply to Peter Norvig

Hi Peter Your article is great and it explains probabilistic and stochastic models quite well. You're absolutely correct to point out that my painting "all ... models" with a broad brush is wrong. I should have phrased it differently.
11-6-1 9:43
Chomsky, it appears to me, is interested in explanatory answers of our natural actions -- questions such as how we recognize a coffee cup on a table full of other items, or how we understand a sentence in a newspaper even though we've never encountered it before. In both the above examples, we now have brilliant technology capable of isolating objects from an image and translating text from one language to the other. However, the point he makes, as I understand it, is that these technologies or their techniques do not offer any insights regarding how humans do the same reflexively and instinctively. It cannot be something as simple as self-training algorithms that are currently employed because even babies are capable of these things (Descartes' Baby Paul Bloom, The Intelligent Eye - Richard Gregory). For example, we have a program, Chinook, that plays the game of checkers perfectly. It can be said that Chinook models checkers perfectly and as checkers players go, it's the best. However, it does not offer any explanation of the kind of computations that occur in the human mind/brain. In fact, it definitely does not represent how humans work. Here's what Science had to say on Chinook: "The complete 10-piece databases contain 39 trillion positions. They are compressed into 237 gigabytes, an average of 154 positions per byte." Checkers Is Solved Jonathan Schaeffer, et al. Science 317, 1518 (2007). We can quite safely conclude that no human thinks like Chinook. The usefulness of Chinook and other programs (such as translation) is not what is under dispute. I agree with Chomsky's objection to usage of "observed use of language". A physicist trying to understand the fundamentals of motion would not spend his time looking out of the window and observing natural forms of motion though that provides much more data than any experiment. He would proceed in very much the way that Stephen Conn details in his comment below. To me it seems that Chomsky and you have different interpretations of success. From your article, it's obvious that you also recognize this in Chomsky's view (A) above and in your own response to it. While Chomsky was wrong about statistical methods having poor applications, to me it seems that he is right about their explanatory principles. It could very well turn out that they will one day explain how the mind works but I'm not holding my breath.
Peter Norvig and 1 more liked this Like Reply
Peter Norvig
Thanks, Vijay, good comments.
2 days ago in reply to Vijay L
I'm still maintaining that statistical and probabilistic models are exactly what is necessary for a good explanation of language. I'm hoping they will be sufficient, but nobody knows for sure. I agree that there are some statistical models that, like Chinook, take advantage of the available hardware: vast, fast, perfectly reliable computer memory. Human hardware is different, so naturally the best program for performing well on certain tasks will be different from the best program for explaining human behavior. Much of the emphasis has been on performing well, but there has been good progress on the explanatory front as well. As I said to Stephen, I'll have to try to write about that next.
3 people liked this. Like Reply
Radunskinaomi87 23 hours ago in reply to Vijay L

Gentlemen, thank you for your very interesting essays, but I am struck by the premise for your thesis, Peter - in precis:
11-6-1 9:43
"As I recall, Chomsky's remarks were in response to an audience member... (not recorded by Technology Review) was something along the lines ..." So, erudition notwithstanding, your arguments are based on non-recorded, unexplained hearsay; a thesis that launches with the expression '... something along the lines' to me says you are going to build an argument with yourself and your own views. Not that there is anything wrong with that ... as Chomsky has no provable part in it, it may look as if you are simply using his name to give veracity to your point of view.
Like Reply
wulfcry 3 days ago

This post says for a great deal anything I wanted to say about this subject it has so much familiarity to my own thinking but in an highly consistent way then I likely could. Its like we merely stud the surface off any matter to understand the underlying instructiveness of any activity coefficient descending slowly and farther in knowledge.
Like Reply
Veryon 3 days ago

I think why and how are on a collision course. I need some statistics to back this theory up.
Like Reply
Reactions
J_Angulo 3 hours ago

From Twitter
RT @kvashee: On Chomsky and the Two Cultures of Statistical Learning http://ow.ly/576tc
kvashee 3 hours ago

From Twitter
On Chomsky and the Two Cultures of Statistical Learning http://ow.ly/576tc
notrance 5 hours ago
28 29
11-6-1 9:43
From Twitter
One more retweet from abzsharma
RT @brainpicker: On Chomsky and the Two Cultures of Statistical Learning Google Director of Research @norvig responds http://j.mp/jNeBrU
katjahofmann 6 hours ago

From Twitter One more retweet from elunca
RT @peter_r_bailey: Fascinating followup post by Peter Norvig to Chomsky's remarks and great comments. http://norvig.com/chomsky.html
peter_r_bailey 7 hours ago

From Twitter
Fascinating followup post by Peter Norvig to Chomsky's remarks and great comments. http://norvig.com/chomsky.html
clipmanF 8 hours ago

From Twitter
On Chomsky and the Two Cultures of Statistical Learning: http://t.co/C3FlS7u via @addthis
alexeieleusis 10 hours ago

From Twitter
On Chomsky and the Two Cultures of Statistical Learning http://j.mp/igcq09
epicgraphic 10 hours ago

From Twitter
Serious reading... RT @peteskomoroch: "On Chomsky and the Two Cultures of Statistical Learning" by @norvig http://bit.ly/jerbor
Show more reactions
29 29

On Chomsky and The Two Cultures of Statistical Learning

Uploaded by

Copyright:

Available Formats

On Chomsky and The Two Cultures of Statistical Learning

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

On Chomsky and The Two Cultures of Statistical Learning

Uploaded by

Copyright:

Available Formats

On Chomsky and the Two Cultures of Statistical Learning