Noam Chomsky, Fundamental Issues in Linguistics (April 2019 at MIT) - Lecture 1

well I'd like to in these two talks I'd like to talk about some foundational issues in particular with the the most important ones I think namely what are the fundamental computational operations that enter into constructing syntactic objects and why these and not other ones it turns out there's quite a lot to say about that since the last time I talked here of any problem some solutions Oh get to this in the course of the discussion as far as I can but I think would be useful to begin with a couple of comments some something more general namely what are we trying to achieve all together and studying language many different ways of looking at it these questions I think are in many ways more important than the particular technical results they raised many questions about what is an authentic genuine explanation a genuine solution and what is a maybe very valuable reorganization of data opposing of problems often posing a solution but not really achieving it these things are worth thinking through I think the basic issues were formulated I think for the first time quite perceptively at the outset of the Scientific Revolution in the 17th century Galileo and his contemporaries who are raising all sorts of questions about received wisdom turned their attention to language as well and they expressed their or an amazement at the miraculous fact that with a couple of dozen it was somehow possible to express an infinite number of thoughts and to find ways to convey to others who have no access to our minds everything that's going on in our minds so in their own words which I rather like quote they were awed by the method by which we were able we are able to express our thoughts the marvelous invention by which using twenty five or thirty sounds we can create the infinite variety of expressions which having nothing themselves in common with what is passing in our minds and nonetheless permit us to express all our secrets and allow us to understand what is not present to consciousness in fact everything we can conceive and the most diverse movements of our soul Galileo himself regarded the alphabet as the most dependence of human inventions because it had these amazing properties and also because as he put it it allowed us to express all the wisdom of the ages and to it contained within it the answers to any questions that we might pose kind of like a universal Turing machine in our terms the court royal grammar and logic actually which was just quoting a paraphrase of Galileo had many insights into logic and linguistics its many ways the basis of modern logic there was a rich tradition that developed exploring what was called rational and universal grammar rational because it was supposed to provide explanations a universal because it was concerned with what was taken to be common to the common human possession of language was seeking explanations including descriptions of even the vernacular which was quite surprising at the time innovative but mainly explanations and Universal trying to find what's common to all languages this tradition went on for a couple of centuries many contributions the last representative of it about a century ago was Otto Jespersen as he put it his concern was how the elements of language come into existence in the mind of a speaker on the basis of finite experience yielding a notion of structure that is definite enough to guide him in framing sentence of his own a crucially free expressions that are typically new to speaker and hearer and also beyond that to find the great principles that underlie the grammars of all languages know that I think it's fair to interject interpret the tradition is metaphoric often vague but I think it's fair to extricate from it the recognition that language is the capacity for language as well as individual languages are possessions of individual persons they're part of a person their shared was recognized throughout the species without significant variation and recognized to be unique to humans and fundamental respects that general program falls within the Natural Sciences within what these days is called the bio linguistic program of course it ran into many difficulties the conceptual difficulties empirical difficulty is the evidence was pretty thin and nobody really understood how to capture the notion yes person's notion of structure in the mind what is that that enables us to develop construct in our minds in many expressions and even to find a way to convey to others what's going on in our mind that's good call the galilean challenge which is still extant well all of this was swept aside in the 20th century by structuralist behaviorist currents which very typically adopted a very different approach to language taking the object of study not to be something internal to the person but some outside thing so maybe a corpus an infinite set of utterances some other external formulation and you see this very clearly if you simply look at the definitions of language that were given through the early 20th century by the leading figures so for example for this is a language is a kind of social contract it's collections he put a collection of word images in the community of speakers for letter Bloomfield languages the utterances that can be made in a particular speech community for Harris it's the distribution of morphemes and a moved to philosophy of language say then why no languages as he put it I'm quoting a fabric of sentences associated with one another and with stimuli by the mechanism of conditions responds elsewhere an infinite set of sentences David Lewis language and languages also took just languages a language is some set of sentences which is infinite that both coin and Lewis crucially argued that it makes sense to talk about an infinite set of sentences but not of a particular way of generating them which is very strange notion if you think about it because these are the lead logicians and philosophers you can't talk about an infinite set in a coherent fashion unless you have some characterization of what's in it and what's not in it otherwise you're not saying anything but the behaviorist the pressure of behaviorist beliefs was so powerful that the idea that there could be a privileged way of generating that infinite set was as wine put it folly Louis but it's something unintelligible but whatever any of these entities are they're outside the individual the tradition was completely forgotten people like yes person the last representatives were literally unknown this good review of this by a historian of linguistic Julia Falk runs through the way yes person was disappeared in the first half of the 20th century of the whole tradition way back also effect this day the even linguistics historical scholarship just pretty thin it doesn't barely recognizes any of the things I've mentioned well so returning to the forgotten tradition by the mid twentieth century it was there was there were clear ways of capturing the concept notion of structure in the mind yes person's concept touring other great mathematicians that established the tools for addressing the galilean challenge something I'm sure are familiar with so yesterday's notion of structure becomes the alcohol the I language the internal generative system finite system that determines an infinite array of hierarchically structured expressions that express thoughts insofar as they can be expressed linguistically and it can be externalized and remoter systems typically though we know not necessarily sound we can call this the basic property of language well to meet the galilean challenge there are several tasks that have to be undertaken the main one of course is to try to determine the internal languages the AI languages of speakers of typologically varied languages a huge task then the question was comes of how a speaker selects a particular expression from the internal I language then how the expression one selected is externalized on the inverse how the external ization is internalized by the hearer the last two tasks are both input/output systems we have grasped how to study those and quite a lot has been learned about it over the years the first of them how the speaker selects a syntactic object out of the infinite array that's a total mystery there's nothing to say about it that's true of voluntary behavior generally so actually some of the two of the leading specialists on the neuroscience of voluntary action Emilio Beattie Robert a Jamie and of a year ago wrote a state-of-the-art article in which they discussed tell what they know about voluntary motion simple not language simple things like lifting your finger you know they said will they put it as they said fancifully that we're beginning to learn about the puppet and the strings but we can't say anything at all about the puppeteer so how you select what you're going to do remains the kind of question that you can't even pose intelligently in the sciences at this stage yours will well the I language keeping to the tradition is a property of the individual and also the species-specific Faculty of language also an internal property something which allows the I language to be acquired and it has to meet a couple of empirical conditions of two conditions which are kind of conflicting the conditions of learnability and the conditions of evolvability so whatever the Faculty of language is it's got to be rich enough so that possessing it a a child can acquire the I language from the scattered and limited data available and it is scattered and limited and it has achieved the internal system which has all of these rich and complex consequences so it has to be that rich but it also has to be simple enough so that it could have evolved and now we can be a little more specific about that because some of the conditions of evolution of language are coming to light and time and the evolution has to meet those empirical conditions well those are the conditions for a genuine explanation if some proposed a descriptive device satisfies these conditions then it's the basis for an explanation for addressing the galilean challenge as it was formulated and developed in the tradition of rational and universal grammar the general explanation is always at the level of ug the theory of the Faculty of language and it has to offer some prospects of satisfying the conditions of learnability and evolve ability that's a pretty austere requirement very austere requirement but it's the right requirement anything short of that is short of actually explaining things it's may be very valuable maybe organizing problems in an interesting way move on from there but still fall short of general explanation that we can now I think grasp somewhat more clearly what actually is a genuine genuine explanation something that was really not possible in earlier stages of linguistic inquiry but again any device that's introduced to account for something unless it can meet these joint these dual conditions is short of explanation may be very valuable so many examples take a concrete example to illustrate about something all come back to later if there's time interesting paper by Boscovich who you all know on the coordinate structure and adjunct island constraints and what he points out is that each of these constraints poses many problems many mysteries but his paper is an effort to try to reduce the mysteries by reducing both constraints to the same constraint using the device of neo David Soni and event semantics which interprets a Junction is a kind of coordination so you can reduce both of the problems to the same problem of coordination and then we still have the mysteries but now a simpler problem one set of mysteries instead of two independent ones and tries to show that the problems then reduce in this way well that's a step forward it leaves the mysteries in a better position and for productive inquiry but it's not an explanation he's quite clear about that and I think if you look over the field that virtually every Chiva one is a partial step forward in this respect there is very few exceptions just barely coming to light which I think can count as genuine explanations they're important in themselves and they're also a kind of a sort of a guideline and to how we should think about proceeding and then they also tell us something about just how far it's possible to go it's not so obvious Ingham much beyond what the kinds of explanations that are now beginning to come to light I'll talk about that well actually the the earliest work and generous grammar tried to meet even more austere conditions the was heavily influenced by work of people like Nelson Goodman and that we've equine who were working on what they called the constructive nominalism no sets very austere just you know Mura logical concepts of a very limited kind that that was too austere at least for the present so that was kind of dropped at least for the present maybe even come back to it some day and attention turned to something else namely the vast range of empirical data from all kinds of languages that was beginning to appear as soon as the first efforts were made to write actual generative grammars turned out that everything was puzzling and complex nothing was understood it's just massive puzzles big change from a few years earlier during the period of structural linguistics it was basically assumed that everything was known everything was solved yet the methods of analysis you could formalize them all that was needed was to just apply them to one or language that turned out to be radically false well the first proposals as you all know were dual they there were there were operations to deal with the problem of compositionality very structured grammar and totally different operations to deal with the phenomenon of dislocation ubiquitous phenomenon transformational grammar both systems were far too complex to meet the long-term goals of genuine explanation that was well understood the general assumption at the time remaining for a long time often open until today is that the principles of compositionality are natural you can expect those something like for a director drummer but the dislocation is a weird property that languages have a kind of imperfection that we have to somehow languages for some reason have this formal languages would never be constructed with that property and that is still a widely held view I think it's exactly the opposite of the truth the opti the opposite I think turns out to be true that more recent work suggests that a dislocation is kind of the null hypothesis it's what's expected on the simplest grounds and it's the most primitive of operations I'll come back to that but let me just take a brief look at the steps that we're taking to reach what I think is this conclusion well in the sixties a phrase structure grammar swirly a limited the phrase structure grammar is far too rich to be contemplated as relevant to describing languages so there's nothing infrastructure the theory of free storage of grammar that prevents you say from having a rule you know VP arrow in CP let's say find Fraser rule doesn't make any sense it was just assumed you just can't do that sort of thing but the right theory has to rule that out as unacceptable and that step was taken by the late 60s basically did the x-bar theory x-bar theory had interesting consequences which weren't really fully appreciated at the time they're obvious in retrospect for one thing x-bar theory notice has no linear order so Japanese and English say roughly mirror images have about the same x-bar theory linear orders on the side somewhere that was a step towards something which i think is now much clearer namely that the surface order of expressions is not strictly speaking part of language it's something else I'll come back to that but if you just look at x-bar theory it's already a step in that direction another thing about x-bar theory is it forces a theory of parameters so Japanese in English say differ and they're going to differ in some choice that is not determined by x-bar theory so some there the speaker and the hearer who's using a linear system of externalization you don't have to use that but if you are using it you're going to have to make a choice as to the order in which you're going to externalize the internal system so x-bar theory itself first is a step towards separating a linear order and other surface organization from what we might think of as poor I language of the I language that's dealing with the galilean challenge constructing the set of linguistically articulated thoughts putting external ization and some medium oi and I think that pictures becoming clearer we'll come back to that well there are also along with the clear progress of x-bar theory there were very serious problems which weren't recognized at the time the main problem is it excludes the possibility of EXO centric instructions everything has to be endo century in x-bar theory and that's just false there are extra centric instructions all over the place simple things like subject predicate or for that matter every case of dislocation without exception all of these give you X eccentric instructions there's no way to describe them at x-bar theory now in order to describe the many artifices were developed so for example if you have a subject predicate construction maybe it was called a teepee or an IP or something like or a VP but that's just stipulation you could just as well call it a NP and this runs all the way through the descriptive apparatus so there was a serious problem not really recognized till the a couple of years ago my own feeling is it's kind of over it's pretty much overcome by labeling theory which tells you in a principled way in terms of minimal search simple computational principle when it is when movement internal merge may take place when it must take place when it need not take place and there are many interesting results and plenty of interesting problems about this a lot of very intriguing material most of which I presume you're familiar with well by the moving up to the 1990s it didn't seem to number of us that it's enough had been learned so it might be possible for the first time to confront the problem of genuine explanation that's what's called the minimalist program pursuing that program if you want thee if you want a genuine explanation you want to start with computational operations which meet the conditions of learner ability and evolve ability well the easiest way to meet the condition of learnability is to say that learner ability 0 it's just innate nothing I'm the easiest way to meet the condition of evolve ability would be to say let's find the computational principle that had to evolve there was no way for it not to have evolved well if you look at those two conditions they're satisfied by the most elementary a computational operation what's been called merge in recent years which incidentally has many problems that we'll come back to but basically just the operation of a binary set formation it it has to be there because the basic property exists ok and that means at least at the very least the simplest operation must exist maybe more complex ones or at least the simplest one so we know that it has to exist had to evolve so it meets the condition of evolve ability oh that leaves the question of just how it happened and what the neurological implication is but whatever the answers to those this is an operation that had to evolve and having evolved it's an 8 so it meets the condition of learnability so if you can reduce something to that you do have a genuine explanation that's as far as it's possible to go ok if it doesn't if you can't go that far it's a description it's not a it's not a genuine explanation again this is a pretty austere requirement but I think it's the one we all have in mind when we're thinking about the goals of our efforts and enquiring into language well so let's I won't give the details because I think you're familiar with them but the simplest computational operation then merged binary set formation a meeting that the no tampering condition least possible computation you don't modify the elements don't add any more structure interesting things to say about this to which I'll come back oh there is a good deal of current literature which tries to show that you can reach this operation in steps that's incoherent you can't have partial binary set formation you can't reach it in steps let's either have it or you don't have it there's nothing simpler again lots of literature about this but it's just beside the point there's actually a recent interesting recent paper by greeny Hoyt births analyzing some of the recent proposals and showing why they don't make any sense but if you think about it they can't make sense the simplest case of merge is going to have at least maybe at most we would like to show but at least two cases that one of them external merge when you're taking separate things and forming this one internal merge when you're taking one thing on something inside it for me the set of those those are at least the two the two simplest possibilities notice there are only one operation there's no no two operations just one operation with two cases much confusion about this and the and the literature but that should be obvious if you think it through well notice that this is this whole program is a program it's not a theory the program is to see how far can we go if we take the simplest possible operation and try to give genuine explanations in terms of it maybe that's impossible maybe have to find more complex operations but in that case it's gonna be necessary to demonstrate how they can be acquired how they can be learned that now they could have evolved and that's not so trivial you can't just say well natural selection does anything I like you know that's not an explanation you can you have to give a real explanation very difficult in biology it's and the biological literature it's pointed out that it's fiendishly difficult eddard phrase to give an account of the evolution of almost any trait even the simplest ones like having blue eyes for example and it's it's not the kind of thing you can hand wave about so either you can try to meet that condition or recognize that you don't have genuine explanations well there have been I think substantial achievements in the last recent years and trying to gain general genuine explanations they do have problems I want to return to the problems later but I'll put them on the shelf for a moment the one one achievement which is not trivial is to unify the two traditional kinds of operations a compositionality and dislocation they are unified once you keep to the simplest the simplest computational operation so far from being an imperfection as was always assumed by me in particular it would take a stipulation the bar dislocation if you have no stipulations at all and you get dislocation furthermore as I mentioned before that's arguably the simplest case of merge actually you can't have only one and not the other because once you have merge of both but if you're looking for one that's more primitive it's probably internal merge the reasons for that are quite straightforward external merge requires enormous search to put two things together that are separate first we have to search the entire lexicon then you have to search everything that's already been constructed and maybe is sitting there somewhere waiting to be merged with internal merge you have almost no search at all so one reason for regarding internal merge this location is more primitive it just doesn't it requires a tiny fraction of the search but there's a good deal more than that there's some interesting suggestions and the literature they're not definitive but they're suggestive so one was some work that was done by Marv Minsky a couple of decades ago he and one of his students just explored what would happen if he took the simplest turing machines smallest number of states smallest number of symbols and just let them run free you see what happens what turned out was kind of interesting and most of them crashed either got into infinite loops or just stopped but the ones that didn't crash all of them gave the successor function no what's the successor function well one thing the successor function is is internal merge so if you take a merge and you have a one member lexicon just run three get the successor that's Minsky's argument at the time was that probably evolution in the course of evolution nature found the simplest thing that's what you'd expect so it found the successor function and that happens to be internal merge external merge there's if you look at other or way down to the level of insects they have a count so an ants they can count the number of steps it's taken it's got a count or maybe a set of counters inside and if you look at just the mathematics of successive counters they kind of tend towards the successor function it doesn't take a big step to move them up to the successor function so from various points of view it seems plausible to think that of the core operations of the most primitive one is actually dislocation contrary to what was always thought and as you get richer constructions you have external merge and is you richer or kinds of languages we plainly have it a natural language it's not just internal merge interesting questions why that probably has to do with argument structure which is uniquely related to external merge come back to that well what's with the unification of internal and external merge compositionality and dislocation what was suggested by x-bar theory as I mentioned before becomes much more clear and explicit so the it seems that the generation of the CI interface sometimes called LF what gets automatically interpreted the linguistically articulated thoughts that's we can call or I language and that just keeps the structure no linear order no other kinds of arrangements so why is there linear order in spoken language incidentally not not strictly in sign language so in sign language which we know to be essentially equivalent to spoken language there's different dimensionality so you can use visual space you can use simultaneous operations facial gestures and motions it's not strictly linear it makes use of the contingencies allowed by the space that's of externalization but speech happens to be linear you have to string words one after another so the if you pick that particular modality of external ization yes you're going to have linear order but does linear order have anything to do with language you know times what you think you want to call language but what it really has to do with is an amalgam of two totally different independent systems one of them internal language the other a particularly sensory motor system which is absolutely nothing to do with language the sensory motor systems are around hundreds of thousands maybe millions of years before language ever appeared they don't seem to have been affected by language at most there's very minor suggestions about slight adaptations that might have taken place for C changes of the alveolar Ridge and click languages there are some very small things but basically the sensory motor system seemed independent of language but since you're if you do externalize the internal system through this filter you're going to get linear order but strictly speaking that's a property of an amalgam of two independent systems and in fact that's true of externalization altogether and notice that the externalization opposes a hard problem you have two completely independent systems they have nothing do with one another you have to match them somehow you can expect that process to be pretty complex and also to be variable you can do it in many different ways also to be easily mutable you can change from one generation to another under slight effects putting all these expectations together with what is a natural expectation I'm not I think increasingly is coming to be imaginable maybe true is that the variety and complexity and buta bility of language is basically a property of externalize ation not a property of language itself and it could turn out to be true all at the moment that the core online which is really unique may not vary from language to language actually that much is pretty much tacitly assumed and essentially all the work on formal semantics and pragmatics it's not assumed to be parametrized from one language to another or to be learned somehow it's just there you know which means if we ever understand it properly it'll be reducible to elementary computations which are just don't worry that's the way the internal system works that should be the goal of inquiry in those directions I should say just as a terminological point what's called formal semantics is actually a form of syntax it's symbolic manipulation technically something becomes semantics when you relate it to the external world and that's a tricky business even things like say even calculus if you think about it events are really mental event mental constructions you can't find them in the outside world and the task of relating what's internal to the external world dealing with questions of reference is no trivial matter we can see a goal for all of this work to try to reduce it to computational operations that do meet the conditions of genuine explanation in a very austere criterion but I think one that's worth keeping in mind well these are all possibilities that I think they're looking increasingly plausible the fuelled may go in that direction be very striking the discovery if it really does well let's go on with genuine explanations one of them is dislocation putting it together with compositionality and notice that that includes automatically the basis for what's called reconstruction you keep to the no tampering condition you automatically get what's called the copy theory of movement that's the basis for the complex properties of reconstruction there's a lot to look into but that's essentially the basis for it you don't need rules of reconstruction there just there that's automatic well of genuine explanations the most interesting case I think is the old principle of structure dependence this was discovered back in the 1950s this is really strange principle of language which had never been noticed namely that the rules and operations of language the ones that yield interpretation of sentences don't pay any attention to linear order they just deal with structures which is extremely puzzling when you think about it because linear order is what you hear it's a hundred percent of what you hear you never hear structure furthermore at least superficially it seems that computations on linear order are simpler than computations on structure from another point of view that turns out to be false but at least superficially that looks right so what it seems it would always seemed extremely puzzling is that the rules that the syntactic rules and the rules that yield semantic interpretations don't pay any attention to a hundred percent of what you hear and to the simplest operations which is a pretty puzzling fact and we now have a simple explanation for it it follows from the simplest computational operation if the entire internal language is based on the computation of the simplest merge operation in its simplest form then you automatically get structure dependence for operations of movement of construal of interpretation of everything else I won't run through examples I assume you're familiar with them but that just seems to be a fact about all constructions and all languages oh it's that if it's correct is a genuine explanation of a fundamental property of language maybe the deepest property of language that the core language just doesn't care about order arrangement it only cares about structure and a child learning language just ignores everything they hear by now there's interesting independent evidence supporting this conclusion studies of language acquisition which have proceeded and very sophisticated ways by now have no gotten down to the point where 30 months old infants have been shown already to observe the principle of structure dependence that's almost no data remember and it's a very abstract principle there's other work earlier work by Steve Krane Nakamura who a lot of evidence three-year-olds have mastered a study recent studies have it down to 30 months if we have better studies which as they keep improving it'll probably be earlier what that means is you just born with it so it meets the condition of learnability namely zero and it has the condition of evolve ability you have to have this a particular operation at least maybe more but at least this one because you do have the basic principle well it's there's also as many of you know neuro linguistic evidence the studies of inspired by Andre Moreau of a group in Milan mousou and others have shown many of you know this that if you present subjects with invented systems of two types the one which correspond to the rules of an actual language that the subjects don't know the other which uses things like linear order you get different kinds of brain activity in the case of say having a negation be the third word in the sentence very trivial operation you get diffuse brain activity if you follow the what looked like more complex rules of actual languages you get activity in the expected language specific areas the brain Broca's area and so that's been by now replicated many times it looks like a pretty solid result there's also psycho linguistic evidence of other kinds the moral Musso experiments were actually suggested by work of neil smith and he on theat simply on a subject they've been working with for many years a young man they call chris extremely limited cognitive capacities almost none but tremendous linguistic capacities he picks up languages like inhale know what that means like a sponge you know the words just picks them up immediately and the neil tried these neil smith tried these same experiments before then the neuro-linguistic the ones were done he just tried it with Chris and turned out when he gave Chris a nonsense language modeled on an actual language he learned it easily like every other language when they gave him the very simple language things like negation being the third-world word he couldn't handle it was just puzzle he can't deal with puzzles that's what inspired the noir linguistic studies I think this is the most interesting discovery so far and the brain sciences related to language it's a direction in which other experimental work could go well looking back at this it seems to be one of these very rare cases where you have converging evidence from every direction leading to the same conclusion that poor eye language just is independent of linear order and other arrangements they have linguistic evidence cycle psycho linguistic evidence or linguistic evidence evolutionary considerations anything you can think about now there's a very curious fact there's a huge literature in computational cognitive science trying to show that somehow this principle can be learned which is a very weird fact if you look at it it's like trying to find a complicated way to disprove the null hypothesis things like that just don't happen in the sciences I mean here you have a the absolute optimal explanation and a huge literature trying to show look there's a very complicated way in which maybe we can reach the same conclusion it's an enterprise that's kind of senseless at the base of it of course when you look at the actual cases never works it's not gonna work if it did work it would be meaningless because always asking the wrong question and I suppose you could show that by the detailed you know statistical analysis with recurrent neural networks and so on of many layers of say the Wall Street Journal you could find the evidence that a child might have used a thirty months old to discover that you have structure dependence you're not gonna find that of course but even though there's literature claiming it but if you did find it it would be completely meaningless of course the only question is why is this the case why is it that in every language and every construction this is the way it works if you could find the way of showing well here's how it might work in this language tells you nothing it's answering the wrong question and furthermore as I say it's trying to find a complicated way to disprove the null hypothesis the whole enterprise is completely senseless it's actually the probably the major effort and computational cognitive science to try to find a basis for some linguistic principle huge literature on new papers still coming up a very strange thing papers trying to show that as it's the way they put it often you can get structure dependence without what's sometimes called an inductive bias for structure defendants but there's no inductive bias it's just the null hypothesis make no assumptions this is what you get there's no bias it's just given so I think interesting question about the many interesting questions about how linguistics is done but one of them is why things like this go on I think we're thinking about well there are other successes but what I'd like to do is turn to problems there were a lot of problems about merge and there's some has solution so one problem is whatever he mentioned EXO cetera constructions so it takes a and B VP let's assume since Dominica is here let's in his honor or some the predicate internal subject you put together or subject Ana an NP and a VP the NPS are often called DP so come back to them I think it's probably a mistake let's just call them noun phrases for the moment you have a noun phrase and a verb phrase you put them together or that gives you the basic theta structure well the noun phrase and the verb phrase have to be independently constructed which means you have to have some kind of workspace something that Jonathan's pointed out years ago you have to have some kind of workspace in which you're constructing these separate things and if you think it through the workspace can proliferate not indefinitely and get larger where you're just doing parallel things and putting them together so it means that the operation merge really oughta be can't revise to become an operation on workspaces not on two elements x and y it's an operation which changes a workspace to another workspace and then the question comes how it does it will know I should say I'm very pleased to be back at a nice low tech institution like my with blackboards and no PowerPoint with no projections and that's which they have in Arizona where here so what we want is some kind of operation that says it's called a capital merge we'll look at its properties which takes two things call them P and Q the guys were gonna merge and a workspace and turns it into some other workspace so what's the other workspace well it's gonna include the set P Q the two guys were putting together in fact let me use a different notation for reasons all mentioned for a workspace is a set but we want to distinguish it from the syntactic objects which are sets so a workspace doesn't merge with something so just to for convenience just use a different notation so the new thing will include the set EQ and a lot of other junk and the next question is what's the other junk in the workspace that turns out to be not a trivial question so what a lot terms on what the answer is so let's take the simplest case okay the entire workspace consists of two elements a colum a and B that's the workspace and suppose we decide to merge them so we get the new workspace which includes the set a B and does it include anything else so for example does it include a and B well if we think about the way recursion generally works it should include a and B so if you're doing say proof theory okay you generating a proof you'd construct the line from axioms and former things and you can go back to that line if you like you can always go back to anything you've produced all ready for the next step but there's a good reason to believe that four organisms and particularly humans doesn't work that way and you can see that if you think what would happen if you did a while this to happen okay well suppose you allow this then you could go on to construct some much bigger object here including a B as a term but it could be of arbitrary complexity you know any kind of complexity alike and then you could take a and merge it with it and get X a but then a would be up here and a would be down there and there are two copies and they would violate every imaginable constraint on movement okay so if you allow this you're going to get total chaos every constraint on this location will be violated no matter how radical you make the the violation well that tells us something it tells us something kind of surprising and I think significant that the kind of recursion that takes place in human language and probably organic systems generally cuts back the number the set of items accessible to computation as narrowly as possible okay let's give it a name and call it resource restriction it looks as though a very general this is merely the first example if you think it through it works for millions of things the same model of refutation eliminates whole set of possible extensions of merge that have been proposed over the years oh come back to examples but you notice what the problem is the problem is that you if you allow the normal kind of recursion you know no constraints and no limits then you're going to find that by legitimate means you can get illegitimate objects oh that has to be barred right you can generate all kind of deviant expressions that's not a problem but you don't want to have legitimate means for generating things that violate every possible condition descriptive condition anything like that is wrong well in this case and it turns out in a great many cases you can bar this outcome simply by limiting the resources that are available now what are the resources the resources are elements that are accessible to the operations so the real condition says limit accessibility okay keep accessibility as small as you can we already have examples like that that we're familiar with one of them is the phrase impenetrability condition okay if you think about what the condition says basically it says when you're generating something you get to a certain unit a phase talk about what it is anything inside the phase is no longer going to be accessible to operations okay that reduces the amount of computational search that's required but it's a way of limiting accessibility okay it says those things down there aren't accessible anymore another example and this may be the only other example is minimal search this is what's often called a third factor property third factor for those of you who are not familiar comes from the just simple description of the elements that enter into computation into learning saying what enters into acquiring a system is three things external data internal structure and basically laws of nature which are independent of the system question so if you're studying a growth of arms let's say humans grow arms not wings partly because of nutrition to the embryo and partly in fact largely because of internal structure just genetic determination and extensively simply because of the way physical laws operate there's only certain ways that organisms can develop other ways just not possible you put these together you account for any kind of growth and development the same is true of language there's external data or whatever it is it's going to determine whether you end up with the call log or English internal structure which at least includes merge or Modell but at least that and in fact anything that can be explained in terms of that it does yield a genuine explanation and then laws of nature what are the laws of nature will language is a computational system unusual fact that's rare organic nature may be unique even aside from counters but anyway that's what language is so among the laws of nature that you would expect would be things like computation elimination of computational complexity doing things as simply as possible several reasons for that one of them actually goes back to Galileo again and one of Galileo's precepts was that nature is simple and it's the task of the scientists to prove it whether it's following objects or flight of birds or the flowers or whatever that's a kind of a prescriptive hypothesis you can't prove it but it's been extraordinarily successful in fact the whole success of the sciences in the last 500 years is based on that and that's a enough reason to assume that it works for us too so reasonable to accept that there's a general point which just has to do with the nature of explanation it's just a fact about explanation that the simpler the assumptions the deeper the explanation that's just kind of logic so there's a lot of convert I should say in the case of language there's another reason to believe it which is unique to language and it has to do with the conditions on evolution of language we don't know very little is known about evolution altogether you know as I said to try to really account for the development of any particular trait is very hard even in simple cases in sort of the you know evolutionary psychology literature everything looks easy you know happened by natural selection why not but when you really try to explain something it turns out to be hard in the case of cognitive development it's uniquely hard because you have no fossil records ok no tape recordings of people we're doing whatever they were doing 100,000 years ago furthermore when you deal with language in particular it's super hard with other organic systems say vision you can you have comparative evidence you can study cats and monkeys which have essentially the same visual system and with them we rightly or wrongly allow ourselves to do invasive experiments so you can stick a neuron into one cell and the striate cortex and see what's happening you can and you learn a lot from that that's how we know about human vision but in language you can't do it because there's no other analogous system it's unique system nothing analogous in the organic world so there's nothing to test okay so it's kind of uniquely hard and nevertheless there's some evidence the evidence at Bob or we can I have a book reviewing it by now there's better elements than what we had in the book there's by now genomic evidence that a humorless ap ins that began to separate roughly two hundred thousand years ago that's when you get the separation of the Sun people in Africa from the rest interestingly they have unique forms of externalization these turn out to be essentially all and only the languages that have complex click systems there are what apparently look like a few suggestions exceptions but they seem to be borrowings or something accidental there's a you know very interesting paper by really really high press on this recently so what we know one thing we know pretty convincingly is that roughly around 200,000 years ago humans began to separate they shared the Faculty of language at the time so there's no known difference between the Faculty of language of the on people and everybody else you know nobody knows any differences group differences of language capacity there happens to be a different form of external ization which suggests and then try greenie goes into this in detail on his article that this particular forms of external ization developed later as a matter of logic the internal system had to be there before you can externalize it that's not debatable but he suggests there's a gap when the system was there roughly 200,000 years ago and began to be externalized in different or somewhat different ways later on when did Homo sapiens appear well here we have reasonably good fossil record which shows that anatomically modern humans emerge appear roughly at that time maybe 250,000 years ago which is essentially nothing in evolutionary time so it looks as though the language emerged pretty much along with Homo sapiens with the Faculty of language intact another kind of evidence comes from the archaeological record which gives you know a lot of information about rich symbolic activity it turns out that almost the entire rich symbolic activity there anybody dug up so far is after the appearance of Homo sapiens well rich symbolic activity has been naturally taken to be an indication of the existence of language also more complex social structures you know burial practices so putting it all together it looks plausible that language emerged suddenly in evolutionary time along with Homo sapiens some whatever change gave rise to Homo sapiens seems to broad language along with it and it apparently hasn't changed since that's independent reasons to believe that whatever is in there is probably very simple along with say yellow land precept and the general principle if you want explanation you want simplicity so it makes good sense from many points of view to assume that the relevant laws of nature here or computing computational complexity so computational efficiency that's what it means to call it a third factor property well one particular case of computational simplicity hi how you doing what are you doing here you graduated years ago go back to tough [Laughter] that's third first of all it's not necessarily the case that's a myth you know there's a myth that neural nets are what do the computation but there's pretty good evidence that that's not true the neural nets just don't have the computational real neurons I'm talking about real neurons real neurons may not be the elements that enter into computation there's by now reasonably strong evidence against that Randy Galveston's book with William King is a very good case where he is strong evidence Randy that if you look at neural nets you simply cannot find the basic elements of essentially Turing machines you can't find the core kind of computational element that yields computational activities is not there in neural nets so what he's arguing is that people who have been looking for neural net accounts of computation are like the traditional a blind guy who's looking under the lamppost for his lost keys because even though we lost them across the street because that's where the light is so yes we know something about neural nets but happens that what we're looking for somewhere else there's also there's a lot more evidence the speed and the speed and scale of computation is way beyond what neural nets are capable of and when others including around these particular proposal is that the computation is actually down at the molecular level I mean do with RNA and so on there are other proposals by not inconsiderable people like Roger Penrose for example that the computation is being done by structures that are internal to neurons which have vastly greater computational capacity there's chemical processes that go on in the brain that aren't captured by neural Nets so the fact that it's been known as far back as Helmholtz that speed of transition of in neurons is just kind of too slow to be doing very much so it looks oh we're going to have to look elsewhere to find the implementation of computational systems but you're there's something there and it's going to be a third factor property something about the brain clearly we talked about this in our book in fact so yes there's surely we're going to try to relate whatever is going on ultimately down to the level of cells you know that's science try to reduce everything okay so that's all third factor if you like but talking about neural nets is kind of like talking about natural selection yeah I thought you brought up neurons something in the brain yeah yeah surely something in the brain is responsible for this not the foot let's say you can amputate your leg and you still have language in state your head you don't so yeah we agree there's something going on in there it's a tricky question it's very hard question even for simple traits not just language a very simple traits as I said you look in technical studies of evolution the phrase that's often used is fiendishly difficult to find the evolutionary basis for even the simplest traits so to think we're going to suddenly find them for languages little misleading there are some interesting ideas so angela Ritchie's book that came out recently MIT first book on the state of the art and neurolinguistics is interesting suggestions about what might be involved in the probably small range change in brain we wiring that led to the appearance of merge or something like it you know closing a certain circuit and the dorsal and ventral connections it's an interesting proposal certainly not going to be trivial you know it's a hard problem yes that would be so no we're wasn't he's always been very disruptive [Laughter] well maybe I'll just kind of end here and go on next time but one principle that we're going to expect for many reasons is computational efficiency minimal searches the most is the strongest form of computational efficiency sir as little as possible and there is a case of restricting accessibility that we're familiar with which I introduced the minimal search that's the case of success of cyclic movement so suppose you've taken a WH phrase and you've moved it up to here and suppose both of these neither of these let's say is blocked by the phase impenetrability condition was that only blocks things here well the next thing that's raised is this one not that one okay we just take that for granted nobody talks about it but if you ask why it's again a minimal search question whatever's selecting the operation argue about what it is that goes back to that mystery I mentioned it's going to take this guy because that's the one that's going to find by minimal search so we have at least two cases already that we're familiar with of limiting accessibility one P I see which is pretty broad the other minimal search which we've just taken for granted and maybe that exhausts it but now I think there's a broader general principle which has just restrict resources that we'll have a lot of effects oh come back to where were examples of that next time there's a temptation at this point to relate restrict resources to something that Ray and I were just talking about the fact that the brain is just slow the done work fest works quite slowly and there's many domains in which that is the case so in many ways the most striking one is vision if you look at the sensory motor system the visual system the cells of the retina are actually responsive to single photons of light okay they're maximally they give you a maximal amount of information the brain doesn't want all that information it's just the way overloaded if it ever got that kind of information inside so whatever the visual system is doing is that for the first step it's doing is throwing out almost all the information that's coming from the retina and apparently every sensory motor system every sensory system is like that the first thing it does is throw away just about everything and try to get down to something limited enough so this slow brain up here can deal with it somehow that looks very much like a general property of which resource limitation of the kind that says don't do ordinary recursion but restrict the resources of which this is some special case all seems to converge kind of plausibly we're very familiar with this in the study of language acquisition so as you all know an infant acquiring of phonetic system it's basically throwing away information it's throwing away tons of information in the first months of life and maybe at about nine months or a year saying the same thing goes on through the rest of language acquisition if you look at something like Charles Yang's general approach the language acquisition where you're just shifting you start the child starts with all possible grammars I languages and then the changes the probability distribution of them as data comes along reducing the probability of things that don't have evidence for them so that they become fundamentally invisible it's also a matter of throwing away lots of information and converging on just a little bit the development of the brain is constantly losing neurons but you don't want all this junk around you want just what you need and the resource limitation fits pretty naturally into that system oh I think I'll stop here and try to come back to more detail examples next time unless somebody else wants to disrupt you

Channel: Abdessamia Bencheikh

Views: 37,807

Rating: 4.9490986 out of 5

Keywords: Noam Chomsky, Linguistics, MIT, Massachusetts Institute of Technology, Language, Generative Linguistics, Chomsky, Syntax, Phonology, Phonetics, Morphology, Grammar, Mental Grammar, اللسانيات, اللغة, علم اللغة, النحو الكلي, النحو, نعوم, تشومسكي

Id: r514RhgISv0

Length: 82min 20sec (4940 seconds)

Published: Thu Jun 27 2019