The Child and the Corpus

EDO SEGAL: Noam, I want you to tell the story the way you'd tell it to a smart fifteen-year-old, the auxiliary-fronting story, the one about how a child forms a question — because it's abstract until it's concrete, and then it's unforgettable. And then, Francis, I'm going to ask you to do something a debater hates. Before you answer it, steelman it. Tell us what the child genuinely has that the corpus does not.

CHOMSKY: Take the simplest thing. From "the man is tall," English makes a yes-no question: "is the man tall?" You move the auxiliary, the "is," to the front. Now, a child confronted with this could form a thousand hypotheses, and most of them fit the simple examples perfectly. The obvious one, the one a general pattern-learner would reach for first, is linear: find the first "is" in the sentence and move it to the front. It works for "the man is tall." It works for thousands of cases. It is simple, it is supported by the data, and it is wrong. Take "the man who is tall is in the room." The linear rule gives you "is the man who tall is in the room?" — gibberish. No child has ever said it. Not once, in any language, in any recorded case. Every child, from the start, reaches past the linear order to the hierarchical structure and moves the "is" of the main clause, not the first "is" it encounters. The child operates on structure before it can tie its shoes.

Here is the force of it. The sentences that would teach a child to prefer the structural rule over the linear one — sentences complex enough to distinguish them — are vanishingly rare in what a child actually hears. Most children probably never hear a single disambiguating example. And yet every child converges, unerringly, on the structural rule, and never even entertains the linear one. Where did the knowledge come from? Not from the data; the data is silent on the point. It came from the child. The child knew, before the evidence, that language is structure-dependent, because that knowledge is not learned. It is the endowment the learning runs on. That is the poverty of the stimulus, and I have never heard an answer to it that did not, on inspection, smuggle the structure back in under another name.

EDO SEGAL: Francis. Steelman first.

BACON: I can do it honestly, because the argument is genuinely good, and a man who cannot praise a good argument from his opponent has no business at this table. What the child has — and I will grant it fully — is a constraint that arrives before the evidence and that fits the world it is about to meet. The child does not waste itself on the infinity of wrong hypotheses; it goes, as if by instinct, to the right family of rules, on evidence too thin to justify the leap. That is real, and it is a thing my own account of induction never adequately faced. I taught that the mind should approach nature emptied of its idols, a clean tablet receiving the instances. The gentleman has shown that an utterly empty tablet would never learn language at all — that some prior shape is the precondition of learning, not its enemy. I concede the steelman entirely. The child brings something. The corpus does not have to, because it is given the answer ten million times over.

Now the answer. You have proven, beautifully, that the child's stimulus is poor and that the child therefore must bring structure. I accept it. But watch what you do next: you say the machine's success is therefore irrelevant, because it learned from a rich stimulus instead of a poor one. And there I stop you, because that is not an argument about knowing. It is an argument about thrift. You have shown the child is more economical than the machine — that it reaches the structure on less evidence. Granted, and it is a marvel. But economy is not the same as truth. The question on the table is not who learned the structure on the least food. It is whether the structure was learned. And by your own admission, both of them end up at the structure. The child gets there on a handful of instances because it was pre-shaped; the machine gets there on a mountain of them because it was not. They arrive at the same competence by different roads. You have proven the roads differ. You have not proven that only one of them arrives.

CHOMSKY: But they do not arrive at the same place, and this is the crux you keep stepping past. The child arrives at the right structure and could not arrive at any other — that is what the impossible-language test shows. The machine arrives at whatever structure is in the data, including structures no human could learn. So they have not reached the same competence at all. They have reached the same outputs on the data you happened to test, which is a completely different thing. Two systems can produce identical surface behavior on every example you check and have utterly different internal organization — one bounded to the humanly possible, one indifferent to it. The child's competence is the right competence; it generalizes in the specific ways human grammar generalizes and in no others. The machine's competence is whatever it was fed. Change the food to an impossible language and the machine follows; the child cannot. You call this thrift. I call it the difference between a system that has the structure and a system that has been shown so many examples of the structure's products that it can fake having it. And the faking is indistinguishable from the having, until you probe the boundary — which is exactly why everyone is fooled.

BACON: Then let me press the one place I think your fortress has a gate. You say the machine would learn an impossible language as readily as a possible one, and you rest enormous weight on it. Have you confirmed this, by instances, in the engine itself? Or is it a deduction from your theory of what the engine must be? Because if it is the latter — if you have reasoned from your premises to a conclusion about the machine's behavior, without going to the machine and trying it — then, my dear sir, you have done the very thing I spent my life condemning. You have argued downward from a dogma instead of upward from the world. Go vex the engine. Train it on your impossible language and your possible one and measure which it fits more easily. I am told this has been tried, and that the results are not so clean as your formulation requires — that the engines do show some bias toward the structures human languages favor. If that is so, your decisive blade is at least somewhat blunted, and the matter is empirical, and we are both, at last, on my ground.

CHOMSKY: The experiments exist, they are recent, and they are mixed and disputed — I will grant you that the question is more open than my sharpest formulations allow, and I have been guilty of stating it more absolutely than the present evidence licenses. But notice what even the strongest pro-machine result would show. Suppose a network does exhibit some bias toward humanly possible structures. That bias is a property of an architecture chosen by engineers and tuned over astronomical data. It is not the specific, biological endowment, present at birth, uniform across the species, maturing on a fixed timetable, that my program set out to characterize. Even in the best case for you, the machine's bias and the child's faculty are different objects that happen to overlap in their outputs. You would have found that a particular engineered system has a particular statistical lean. I would still be holding the original question, untouched: how does the human child, demonstrably, do it on almost nothing?

EDO SEGAL: Mark this — because something just happened that I didn't expect this early. Francis pushed Noam onto Noam's own turf, the demand for instances, and Noam gave ground: the impossible-language claim is more open than his rhetoric. And Noam pushed Francis onto Francis's deepest commitment — go to the world, don't deduce — and Francis took it as a victory rather than a wound. That's the strange topology of tonight. Each of you keeps winning by invoking the other man's principle. Hold there. The next round is the one where you stop agreeing about the machine and start agreeing about something darker — that the machine, whatever it knows, deceives. Francis built a whole catalogue of how minds fool themselves. We're going to find out if it's a catalogue of how the machine fools us.