Pullum, G.K. & Scholz BC (2002) Empirical assessment of stimulus poverty arguments. Linguistic Review 19: 9-50
This article examines a type of argument for linguistic nativism that takes the following form: (i) a fact about some natural language is exhibited that al- legedly could not be learned from experience without access to a certain kind of (positive) data; (ii) it is claimed that data of the type in question are not found in normal linguistic experience; hence (iii) it is concluded that people cannot be learning the language from mere exposure to language use. We ana- lyze the components of this sort of argument carefully, and examine four exem- plars, none of which hold up. We conclude that linguists have some additional work to do if they wish to sustain their claims about having provided support for linguistic nativism, and we offer some reasons for thinking that the relevant kind of future work on this issue is likely to further undermine the linguistic nativist position.
I found Pullum & Scholz's (2002) paper quite difficult to read through. I am assuming because they target specific studies and go into a lot of details to break down the wrong-doings and argument analysis of a given study; and to do this needs a good grasp of the material at hand. For this reason I will work through their argument against Gordon who’s study was also featured in Pinker’s reading.
ReplyDeleteGordon studied language acquisition of children between the ages of 3 to 5. He wanted to show that “irregular [ex: mice] but not regular plurals [ex: rats] can be used as the non-head (non-rightmost) part of a compound”. In this way it is more likely to accept the grammatical structure of mice-infested versus rats-infested. Pullum and Scholz’s give a couple of counter examples to Gordon’s rule, including a study by Sampson on regular “plural nouns as non-head elements in compounds” (ex: US English will say a drug problem, UK English will say a drugs problem). In this way, the UK English sentence provides counter evidence to Gordon’s claim of universal grammar. Although the paper highlights the wrong-doings of Gordon’s experiment, they do however mention that this is the right path towards figuring out the credibility of the argument from poverty of the stimulus, and inaccessibility claims. (which I find confusing because at the beginning it was mentioned that “we have some concerns about Gordon’s experimental design” and they don’t quite get into the gist of it (p.24))
If they are in fact saying that Gordon is using the right method, I am not understanding their argument anymore.. Pullum and Scholz showed counter examples of the compound results of Gordon’s study, are they trying to say that something needs to account for these as exceptions?
Actually, it's not the details that matter here, but the overall argument:
DeletePullum points out that it's not entirely true that violations of UG are never heard or written. Mistakes sometimes happen.
But not enough to count as negative examples, from which the rules can be learned by trial and error.
What Chomsky said about the poverty of the stimulus remains true: The child hears only positive examples. Everything (almost) that it hears is UG-compliant. And just about everything it says is UG compliant too. No corrections (except of conventional grammar.)
So the child "knows" UG already.
I agree with the thrust of this article. Though I am willing to be convinced, in the current situation, I find no convincing evidence that Universal Grammar exists. The poverty of stimulus is no true poverty, and it rests on the fallacious assumption that we cannot learn from negative evidence. There are many, many examples of things that we have rarely or never seen people doing – kicking someone instead of shaking their hand as greeting; intentionally poking a hole through their paper as they write; squeezing next to someone on the same seat on the bus. However, we apparently learn by taking the conservative tack – if we haven’t observed it, we don’t do it. But we are aware of exceptions – say, if the bus is so crowded that it would actually make everyone more comfortable were you to sit three to a bench. But nobody is arguing that “politeness” is innate – we just apply broad patterns such as “bus etiquette” and “space maximization”, and successfully resolve conflicts between the two when they occur. Why language is a special area of debate probably has to do with its apparent complexity (though it is worth noting that, although we take politeness for granted, it is no simple matter to learn the social cues of a given society). But, as I mention in my previous skywriting, due to the fact that the plasticity of the human brain decreases dramatically with age, it makes perfect sense that we are unable to understand how it is that children learn language so effectively. In line with current theories of synaptic plasticity, second language learners can also learn these subtle patterns, albeit more slowly and with the almost guaranteed risk of stopping short of native-like competence. So we are always able to pick up these subtle patterns, and children do so faster. Nowhere does innateness or a priori knowledge of UG become a necessity. Unlike Pullum, I do lean towards thinking that there is no innate UG. However, I agree that the methodological tools proposed must be employed if we are to begin a productive discussion of the matter. But the onus is on proving something’s existence, and for now, this has not been done for UG.
ReplyDeleteHi Dia,
DeleteI'm just wondering how no innate UG would be able to explain how languages that have never come in contact still exhibit the same syntactic patterns?
One example I'm thinking of is ergative-absolutive languages: most Austronesian languages and the Mayan family of languages are ergative-absolutive, but they've never come in contact with each other. Do you really think it's possible that both of these large language families spontaneously happened to develop the same system to mark subjects and objects? Isn't it more likely there is one common syntactic structure to all humans that can be realized two ways: either ergative or nominative?
About the critical period: do you know if there are any studies that map the rate of language learning relative to other skills at that age? things like social cognition also makes relatively large leaps during that age, but even then, it seems that more language is learned than any other skill during this time period. This is consistant consistent with the idea of early language development being a specialization of an dedicated structure.
DeleteOf course this is all speculation and I do heavily agree with the authors that more empirical tests are necessary. I don't actually know how you would go about measuring how learning takes place in a way where you could compare different domains
Hi Dia,
DeleteI’m just a little confused about what you mean by "[poverty of the stimulus argument] rests on the fallacious assumption that we cannot learn from negative evidence." Do you mean that APS assumes we are unable to learn from negative evidence (as in we are physically incapable), or that UG says that we DO NOT because of lack of enough instances of corrective feedback, or at least that we do not receive negative feedback any more as often as positive feedback? While the former seems implausible, I don’t understand your argument for the latter, either. Assuming that it really is true we do not receive enough negative feedback to learn through trial-and-error type reinforcement, or that we do not receive it any more often than positive feedback for the same types of sentences, then wouldn’t that automatically mean that we really can’t learn from negative evidence?
Unless you really mean to say that it is possible to learn WITHOUT negative feedback (as in with only positive feedback)? If that were the case, although I would agree that ‘politeness’ is definitely not innate, as there are many people who clearly lack it…and children often do things that are impolite…like tugging on strangers, taking things they want that don’t belong to them…etc; however, I would also say that the example doesn’t really show that politeness is learned without negative feedback. Children DO receive negative feedback from parents when they tell them not to point because it’s rude, or that “give me” is not appropriate and that they need to say please. So, although politeness is still complex, there is still, arguably, enough feedback for trial-and-error learning (unlike language). While I agree that Pullum has showed that the current evidence does not really support UG and that more research needs to be done, I really don’t think that we can discount it entirely…
Last thought (and super out there)…
Universal Grammar is specific to the rules of language. You posit that language is completely learned. Well, Grammatical rules have a lot of similarities mathematical rules (there is recursion eg. in the equation [(1 + 5) x (3 – 1)]-2). I would think that if UG did not exist, and instead of language, children were only exposed to mathematical equations, that they might be able to be math geniuses by age 3…because the syntactic rules of math would all be learnt.
I am quite convinced that if " instead of language, children were only exposed to mathematical equations, that they might be able to be math geniuses by age 3" is a true statement. If mathematics were absolutely critical to communicate, and that was how you could make your desires and needs understood by others, then most certainly children would be "math geniuses" to the extent that was required by their society! We are amazingly social creatures, and ordinary humans will learn any skills that allow them to fit in.
DeleteNegative EVIDENCE, not feedback. As in, learning that you cannot say things like "Where did John bring the flowers and?" because you never hear it. Which is completely possible to do, but PS says we need UG to constrain us from saying things like that / that if UG didn't exist we would all be running around saying random phrases in rogue grammar. Which is ridiculous.
Dia, could you explain how you could learn what's in a category and what's not in a category without the negative evidence (of what's not in the category)?. I think you are mistaken (at least for something of the complexity of UG (as opposed to "go on green and stop on red").
Delete“In the absence of any answers, or any research that is even directed at trying to develop answers, we conclude that the APS instance that Chomsky has used to support the APS for over twenty-five years has no empirical support and offers no support for the existence of innately primed learning.”
ReplyDeleteIn the Youtube lecture by Professor Harnad, Chomsky’s Poverty of Stimulus argument was stated as the absence of negative evidence. Pullum and Scholz seem to pull apart a number of studies and demonstrate that these studies are false, and that there is indeed evidence in the language environment of a child that provides instances for learning specific syntactical structures. Even if this were the case, that there is indeed enough evidence for acquisition of these structures, it still does not account for why children don’t make some mistakes. The arguments by Pullum and Scholz do not seem to align with the definition in the Youtube lecture. Even if there were enough evidence for the child to learn the language, and the poverty of stimulus argument were really refuted, I don’t see how linguistic nativism could be excluded.
I think that Pullum and Scholz are trying to do in their article is to ask some questions regarding the specificity of the poverty of the stimulus argument. They overtly state that they are not defending empiricism. They claim to be “agnostic at this point, perfectly prepared to accept that perhaps at some time in the future data-driven algorithmic techniques for language learning will hit a serious roadblock, or that credible evidence for innate task-specific language acquisition mechanisms will turn up.” I think the crux of their argument is that the poverty of the stimulus is too vague and often interpreted differently by different people. They wish to see this argument “develop some explicit and quantitative answers to such questions if they wish to support their claim”.
DeleteP & S do not give evidence that there is negative evidence for the language learning child. All that childreb hear and say is UG-compliant (positive evidence).
Delete"The literature on unsupervised learning in computational linguistics (Brent 1993 and Schütze 1995 are two examples) is demonstrating that algorithms can be learned much more from text than most linguists would have thought possible."
ReplyDeleteThis statement was mostly puzzling for me. The authors claimed that new arguments were weakening PSA by "revealing that more can be learned in a data-driven way than was previously though", and they gave the unsupervised learning as an example. Unless there is something about the definition of "data-driven learning" that I am not getting, I was under the assumption that most of language acquisition relied on data-driven learning.
As Pinker outlined in his article, children that were "discovered in the woods" revealed to be mute. In addition, children need to receive enough data (hear enough language or receive enough feedback) in order to switch "ON" or "OFF" some crucial parameters. Thus, although one could make the case that humans have innate cognitive capacities to pick up language in a rapid fashion, it is still "data-driven".
The authors were discussing the "unsupervised learning" example, yet they are claiming that algorithms can be learned much more from text. Again, unless there is something about "unsupervised learning" that I am not understanding, I would have assumed that the issue of acquiring a spoken language is much more "primal" than written language. Why make the case on PSA about learning from text when learning from others' spoken language is much more important (in my opinion). Besides, I don't see how their "unsupervised learning" is any different than positive evidence. Children are simply learning language by hearing other people conversing rather than direct and corrective information.
“I was under the assumption that most of language acquisition relied on data-driven learning”
DeleteI believe that one of the major differences between language acquisition and data-driven learning is that language acquisition only allows for certain possibilities of outputs whereas data-driven learning can allow for any possible output allowed by the data. For example, take these two sentences.
1) The cat that is on the mat is black
2) John is in the store that is on the hill
Now take the following two questions that can be created from the sentences and take note of which “is” is brought to the beginning of the question:
1) Is the cat that is on the mat black?
2) Is John in the store that is on the hill?
Notice that in question 1, the second “is” is brought to the beginning of the question, where in the second question, the first “is” is brought to the beginning of the question. If language acquisition was purely data learning, then either “is” can be brought to the front to make a question. However, only one specific “is” can move to the front to make the question. For example, if you moved the first “is” instead of the second “is” in question 1, then the question would become ungrammatical and would read: *Is the cat that on the mat is black?
If language acquisition says that you learn which “is” you can move without anyone overtly telling you because you have innate constraints. Obviously, the input you receive influences your output because different people can speak different languages. When something is data driven, both the constraints are driven by the input. In language acquisition, the constraints are innate.
The poverty of the stimulus argues in favour of language acquisition as it states that the input children receive in first language acquisition lacks certain evidence for certain rules, but the children can learn those rules anyways. Pullum and Sholz are simply asking nativist language acquisition to be more specific and “develop some explicit and quantitative answers to such questions if they wish to support their claim”.
Florence, unsupervised learning means no error-correction; hence positive evidence only. The fact that in some cases some kinds of rules can be learned from statistical regularities in written text without error correction certainly does not mean that the language learning child can learn UG from what it speaks and hears.
DeleteI agree with the authors in so far as they point out that it is hard to make assertions on the type of positive evidence children are exposed to within their environment. However, I do think they are missing the point about the poverty of the stimulus argument. When they set up the modus ponens argument for APS, they seem to assert that either the child has innate language abilities or they learn from their environment. The crucial aspect of the Chomsky's idea that has been further developed, however, is that the child needs both. The innate, built-in component is what allows a child to hear language in their environment and learn it efficiently - it is not one or the other as Pullum and Scholtz seem to suggest. Furthermore, the most convincing argument for Universal Grammar or the fact that there are language specific components within the brain that are separate from other learning mechanisms is the multitude of lesion studies and disorders that exist which only effect language and not other cognitive abilities. In my mmmind, the pooooverty of the stimulus argument is meant to compliment this assertion. Insofar it seems a bit trivial to point out problems within the poverty of the stimulus argument that are so highly specific. Yes, agree with the authors that a more standardized, consistent and thorough approach needs too implimented. More emphasis should be placed on looking into language acquisition in other parts of the world as well. In my mind, pointing out flaws is four highly specific examples of poverty of the stimulus is not sufficient to doubt the hypothesis entirely.
ReplyDelete“When they set up the modus ponens argument for APS, they seem to assert that either the child has innate language abilities or they learn from their environment.”
DeleteThe authors do admit that the strict dichotomy is probably a false one, but they keep it to simplify the argument (17-8). To be sure, it is not simply a dilemma between innateness and experience, but between domain-general versus domain-specific learning mechanisms. You can learn to play any sport or play any instrument, but you don’t need to, so it’s unlikely that you are born with a piano-learning program. Hence, the underlying mechanism of all these learnings is probably the same general-purpose mechanism. As I understand it, the question is whether language is just another instance of this general learning ability or if language is learned differently through its own mechanism relying on innate structures. The assumption is that if our grammars are constrained in such a way that is not explainable in terms of experience, then we have evidence for innate language-specific structures. If experience can explain it, then out of parsimony, we cannot posit these innate structures.
“Furthermore, the most convincing argument for Universal Grammar or the fact that there are language specific components within the brain that are separate from other learning mechanisms is the multitude of lesion studies and disorders that exist which only effect language and not other cognitive abilities.”
I think that lesion studies and disorders do support the view that some parts of the brain and some processes become highly specialized for language. This specialization is compatible with the innateness of a language-specific learning mechanism, but to use this as evidence amounts to a confirmation bias. The fact is that, with very few exceptions, everybody is born into a situation with language; so we have little to no evidence of what happens when a human being is brought up without language. But of course, to go ahead and test that would amount to abuse, so we have to count on logical arguments like the APS.
I agree with your last point Daliah that four highly specific flaws are not enough to doubt the hypothesis entirely and I equally agree with the authors a more standardized and thorough approach needs to be investigated. Although I wasn’t that familiar to UG and the Poverty of the stimulus argument prior to reading these texts, one thing does intrigue me, if innate UG and environment are important then why is it that a feral child, who has been isolated from human contact or locked up in an attic not able to learn a language once they are exposed to the proper environment? Genie is an example that comes to mind, Genie was locked up in an attic and was only able to later learn language but never grasped it completely and only talked in a very basic way. If UG is integrated prior to birth, why can’t a feral child learn to speak when the appropriate environment is later present? Genie is not the best example since she was only freed at age 13, which means she is past the synaptic and metabolic peaks Pinker described in his article to be around age four. But if a child with no prior experience to language was found at age four, at the peak of the brain metabolic activity and with 50% more synapses than adults, I am not convinced that he would be able to learn language to the same degree as a child exposed from birth, because he is missing 4 crucial years of language learning. Yet it seems to me that if UG exists than with the proper environmental setting, every child should be able to learn to speak because it is innate, even if this is at age 4.
DeleteLike Kevin says, “we have little to no evidence of what happens when a human being is brought up without language.” And experimentation on this would be unthinkable.
Daliah, no one has ever found a ÙG-specific lesion disorder.
DeleteCatherine, I'm pretty sure a child that was language-deprived till four or five could still learn it.
Why can't we have both innate and data-driven learning? It seems to me that we might need both in order to communicate. Although, my gut feeling tells me that data-driven learning is the answer.
ReplyDeleteData-driven learning is very obvious to see, I mean new words are created everyday. 20 years ago, "email" was not a word, nor the verbal form "to email", "emailing"--somehow we have managed to create new words, put them in different forms and use it. We hear other people use it, figure out what it means through the context, then use it ourselves. And so the usage spreads. Innate language is not enough to support this, at least I don't think it is.
Support for innate language, to me, is largely the facts that a lot of very different languages have syntax in common (then again, maybe at one time or another they could've somehow been shared) and that there are language specific deficits that result from brain damage (as mentioned above).
My first idea was that I don’t see how anything purely linguistic (ex: universal constraints on syntactic structures such as subjacency for example…) could be innate and thus encoded in our genes. Linguistic constraints are probably determined by our limited capacities of memory, attention, etc…What I mean is that there could be no purely linguistic innate capacities (ex: movement out of certain phrases is impossible), but simply innate memory, attention,…capacities (ex: if it were possible to perform certain movements out of certain phrases, we would not have the capacity to keep track of what comes from where) that would put constraints on how syntactic structures can be built. Universality would still be explained this way. To show that this could explain linguistic specific deficits too, we would need to see on which non-linguistic capacity a particular linguistic constraint relies and see whether when the brain area responsible for this non-linguistic capacity is damaged, the corresponding specific linguistic structure is impaired too. But maybe I am contradicting myself here: how could we be able to form ungrammatical sentences if we did not have the required memory capacities to formulate them?...
DeleteIf I interpreted it correctly, the arguments for UG vs. data-driven learning don't have to do with language itself, but rather the rules of grammar. I'm sure both sides concede that the there are both innate and learned parts of language...clearly there is something innate if the capacity for language is truly a human trait (innate to humans) and learning probably has a huge impact on vocabulary of all things. What UG argues is that the grammatical rules that underly language are innate (like you mentioned about syntax) because there is not enough negative feedback to tell children that certain syntactical structures aren't appropriate and yet they rarely make these mistakes.
DeleteI believe the authors were distinguishing 'innate learning' and 'data-driven learning' as completely separate for the sake of argument (and perhaps in the hopes of ridiculing the idea that such a process as innate learning could exist). It is definitely hard to grasp that there could possibly be two mutually exclusive (kid sib: either one or the other) learning processes in our complex and plastic (easily molded) brain. It seems to me that for any specific inborn capacity to be activated, you need the genetic predisposition as well as environmental stimuli. Further performance of the capacity would, again, be influenced by both. I believe that both worlds rely on one another no matter what cognitive capacity we're talking about.
DeleteUG is unlearnable by the L1 child for lack of negative data. The parameters of UG are learnable, and learned.
DeleteThere is no such thing as "innate learning." If something is learned, it's learned, not innate, though the ability to learn is certainly innate. Something can also be partly innate and partly learned.
All learning is "data-driven." There is poverty of the stimulus when the data are not enough to learn from (e.g., lack of negative data, errors, or error-correction, in supervised/reinforcement learning).
The word “Innate” is used frequently in this article, and it’s a word that I feel could use some clarification, especially in the specific context of this paper, which discusses innateness so deeply. Does “innate” here mean ‘born with’? Does it mean ‘not acquired through learning’? does it mean ‘genetically transmitted through generations’? There are many possible interpretations of “innate” (which is defined in the dictionary as “existing in one from birth”, “inherent in the essential character of something”, or “originating in or arising from the intellect or the constitution of the mind, rather than learned through experience”). In this paper the authors seem to use “innate” to mean something along the lines of ‘not acquired through sensorimotor learning’. However, this seems to be something that should have been aptly described and defined in the context of this paper, as it so central to the paper’s arguments, and there are so many possible connotations to the word within cognitive science.
ReplyDeleteIt is so true that the word 'innate' might be misleading. If it does mean 'existing in one from birth' or 'not acquired through sensorimotor learning, then there are many things that are innate, such as our physiology and emotions. What I think is interesting to point out is that although something may be 'innate', that does not mean that it cannot differ amongst individuals anyway. There can be variation on innateness, can't there?
DeleteThere can be variation in innate traits, but there is none for UG, except parameter-settings -- but those are learned...
DeleteThe authors have made a case here that empirical investigations of the poverty of the stimulus argument fall flat when closely examined. They bring up four specific linguistic rules that historically have been posited as instances where children (1) know the correct grammar and (2) have not heard enough examples of the case to determine the correct grammar by way of example or trial and error. In each case they argue that children would in fact have a chance to hear plenty of examples of these rules, calling into question the established empirical evidence for the poverty of the stimulus argument. Of course this doesn't mean there is no innate learning, but it means that more research must be done to verify the poverty of the stimulus argument if it is going to be talked about as a sure thing.
ReplyDelete“ We question whether children learn what transformational generative syntacticians think they learn. If the types of constituent that can occur as complements to subordinating words like that and than are learned piecemeal from positive examples, then the pattern ëmore than + finite VPí will be learned (after examples like They wanted more than was available), but ëV + that + finite VPí will not be learned, because no examples of that sort will ever be encountered. On the other hand, ëN + that + finite VPí will be encountered, and so relative clauses of the type (things that were available) will be learned. In sum, piecemeal learning of constructions here would do better than rapid generalizations concerning the scope of wh-movement. “
ReplyDeleteThis nugget I think challenges one of the core assumptions of UG proponents. It is presupposed that kids learn these general grammatical rules that linguists can formulate as covering an enormous range of sentences, but that may just be begging the question. If nativists are implying that we are born with the ability to detect grammar, assuming that kids are able to generalize from a set of sentences to highly abstract grammar rules is circular. As the author’s suggest, an alternative could be that the kids are doing local abstractions of specific grammar rules from individual sentences, rather than the whole grammar emerging out of them. This method of learning would be more in line with data-driven theory, and it is just as plausible that children learn a large series of individual grammar rules in their critical language period rather than a few fundamental ones. Given that they are exposed to millions of sentences, it would be hard to disentangle these two approaches.
More generally, I think the author’s demands for greater empirical testing is very relevant. Ideas can easily turn to dogma in fields like these where elegant hypotheses make intuitive sense for solving intractable problems.
In my comment to Pinker's Language Acquisition paper, I tried coming up with reasonable learning solutions to linguistic phenomena that apparently lacked explanation without positing a UG that would do the job. As I read Pinker's books, I had the same feeling creeping in of assuming he was sometimes jumping the gun on damning evidence.
ReplyDeleteHowever, there are many cases that I had to take for granted and simply accept (I only have so much time and curiosity!). My underlying assumption was that language learning was very similar to other basic pattern learning, associations made due to the correlational structure of the world. The process would seem like a lengthy process of associations embedded in other associations. This isn't a claim that there is no innate language learning ability, since there must be since children have to normally be forced to learn most other things, while language, aside from seeming like a nuclear weapon type device (as Stevan said), seems to be learned and created (see pidgin and creole) automatically. It's the drive for language learning which seems innate. I believe the UG, as an innate deep structure, is still up for debate. Clearly we innately learn language, but is the mechanism language specific? This passage resonated with me:
“To posit learning from experience for facts of complement type selection is
no more than would be assumed for learning any other structure of parts recombining
to make particular types of wholes. We learn at a very young age
that houses contain rooms, and rooms are of different types such as kitchens
and bedrooms; that nearly all houses have both kitchens and bedrooms; that
cookers and sinks are found in kitchens, and sinks have taps, and faucets produce
water; that dressers are found in bedrooms, and dressers have drawers,
and drawers have knobs; and so on. Our experience does not establish for us
definitively that dressers with faucets are impossible, or that bedrooms do not
contain cookers; yet we come to believe these things too defeasibly (perhaps
some people might have a cooker in their bedroom), and in a way that is dependent
on our environment (we would not have learned the same generalizations
growing up in a rural African village), but without explicit negative evidence."
The analogies we draw between similar but distinct physical contexts seems to be similar to the analogies we would draw between similar types with different words in language. This angle on acquisition seems to be compatible with present models of our left hippocampus as a cognitive linguistic map. Though specialized for the linguistic context, it would be reasonable to assume it uses a similar mechanism as the right hippocampus which maps out the physical world since we can only assume most adaptations begin as exaptations (co-opting a structure or function). The interplay of these two might prove to be how the symbols (words) seem to make us dance.
The poverty of the stimulus argument goes like this: there are things that kids know about language that they can’t have learned from experience. So kids must know those things about language innately, its hardwired.
ReplyDeleteThe poverty of the stimulus argument is really really popular. In their paper, Pullman and Scholz don’t try to refute “linguistic nativism” (the view that some of our knowledge about language is innate), but rather try to show that no one has provided good experimental evidence for the Argument for the Poverty of the Stimulus (APS). After reading Pullman and Scholz’s paper, I am left with this, APS is an interesting idea, but no one has provided good evidence for it. Not even close!
Pullman and Scholz set out a pretty nifty framework for what type of evidence is needed to support the APS. It looks like this. (1) a description of a property of language that kids know (this is called the “acquirendum”) (2) a description of what types of experience a kid would need to learn the acquirendum – aka what sentences they would need to hear an adult say to (this is the “lacuna”) (3) explain why the kid would need to hear the lacuna in order to learn the acquirendum (4) evidence that the kid never hears the lacuna (5) proof that kids actually know the acquirendum.
Then Pullman and Scholz go through a whole bunch of studies that supposedly provided evidence for APS and show that they actually don’t. Often, their main criticism relates to the 4th element of the nifty framework: they argue that kids acutally get the experience that they supposedly need to learn the acquirendum.
To elaborate on this some more, Pullam and Scholtz make an argument against APS because there is not enough empirical evidence to support it. This is a legitimate concern in the grand scheme of scientific theory but it only made me think of the Big Bang theory. There are obviously no records of it so we cannot guarantee but APS like the Big Bang theory left traces of its existence. The former leaves children at 3-5 having grammatical structures approaching or equivalent to adult sophistication. There is no way to test it really but it is reasonable to infer from the traces UG and APS leave behind.
DeleteOther than being extremely lengthy, I liked this article, and I appreciated its critiques. It was very interesting to learn about the aspects of linguistics which completely lack empirical evidence. And yet still, grandiose claims are being made (assuming this critique/article is accurate). From "science" I hope for more.
ReplyDeleteIn attacking the claims of the linguists, I would have preferred the authors addressing another one of the properties of the child's environment. They considered the APS as the strongest argument the linguistic nativists have, but personally I thought the "incompleteness", "positivity" and "degeneracy" make a much larger contribution to the overall claims of the linguists.
Overall, I still buy the linguistic nativist ideas. UG just makes sense at this point, although admittedly I don't know of all the other theories which could potentially human language as well out there.
On a different note, why had Fodor described "the APS as 'the existence proof for the possibility of cognitive science"? This was quickly mentioned at the end of the article, and I'm really curious why? (APS = the argument from the poverty of stimulus.) I understand APS as consequential for cognitive science, but certainly not existential. Why would cognitive science be so dependent on APS, afterall linguistics is just one facet of cognitive science.
The APS and "positivity" are the same thing -- or rather the lack of negative evidence is the heart of the APS.
DeleteNo, APS is not the heart of cogsci, nor essential for it. Chomsky's nativism is evidence-based. Fodor's is theory-based. (He thinks not just UG, but all "concepts" are innate. But all evidence goes against that. Categories are learned, and there's plenty of positive and negative evidence.)
Chomsky did launch cogsci, though, with his critique of Skinner's behaviorist theory of how language is learned. But Skinner would have failed to explain language even if there had not been the APS.
(Fodor, in his solipsistic theoretical slumber, imagines that "learning" means "Lockean association of ideas." He does not seem to be able to grasp modern learning models, whether computational ones or neural nets. He thinks it's all Locke and Skinner. I don't think he even distinguishes Skinner from Pavlov...)
Chomsky, N. (1959). A review of BF Skinner's Verbal Behavior. Language, 35(1), 26-58.
DeletePullum and Scholz discuss four distinct cases where they believe the poverty of the stimulus argument doesn’t hold up. I feel like their argument doesn’t really make as big of a splash in terms of derailing poverty of the stimulus as they are under the impression it does. First of all, I thought (possibly incorrectly) that the poverty of the stimulus argument rested more heavily on the claim that first language learners are exposed to little to no negative feedback (ie. they don’t get corrected very often) than the claim that they don’t have access to enough positive data (ie. examples of proper use of an aspect of language). Yet, Pullum and Schulz seem to focus on the presence of positive data in their defence of data driven learning. (Can’t language acquisition be a combination of data driven learning and an innate device?) I guess what I’m trying to say is I didn’t read this article and think, “Oh wow, universal grammar must not exist because in four cases there actually is enough positive data for first language learners to acquire an aspect of language without some kind of innate knowledge.” Maybe Pullum and Scholz have pointed out areas where Universal Grammar doesn’t hold up and needs to be modified but I don’t think they’ve disproved poverty of the stimulus and contrary to what they propose I don’t think we need to urgently scramble to come up with another example of a situation where there isn’t enough positive data for a learner to acquire an aspect of language. At a certain point firing examples back and forth becomes a little redundant. Identifying a few little flaws in Chomsky’s theory isn’t enough to make the whole thing collapse. Once you’ve proved that there’s enough positive feedback to acquire one aspect of language without an innate device you have to start proving there’s enough positive feedback for another aspect and so on and so on.
ReplyDelete“Although this article will no doubt be misread as a defense of empiricism, that is not an accurate characterization of it. We are agnostic at this point, perfectly prepared to accept that perhaps at some time in the future data-driven algorithmic techniques for language learning will hit a serious roadblock, or that credible evidence for innate task-specific language acquisition mechanisms will turn up.”
ReplyDeleteI personally found the Pullum and Scholz paper to be overly complicated, but this is perhaps due to the fact that I have never taken a course in linguistics or the psychology of language. That being said, I did appreciate that they thoroughly evaluated the literature on the empirical assessment of stimulus poverty arguments. I thought that the quote mentioned above was important because often times when researchers critique an existing hypothesis or theory, it is seen as a complete revocation of that hypothesis or theory, but Pullum and Scholz are willing to accept the poverty of the stimulus argument if data-driven algorithmic techniques fail to come to progress. Also, many researchers in psychology, cognitive science, and linguistics seem to blindly accept the poverty of the stimulus argument, so at least Pullum and Scholz have actually taken the time to do a thorough evaluation of its credibility. The fact that there are a lack of empirical studies that illustrate the validity of the poverty of the stimulus argument is troublesome, so I think, if anything, this paper will urge cognitive scientists to produce empirical evidence in support of it. In addition, I felt that Pullum and Scholz were cautious when they mention that they are in no way eliminating the possibility of a domain-specific innate mechanism specialized for first language acquisition. I found this important because I personally still believe in the poverty of the stimulus argument, even considering the ideas put forth in this article.
Although widely accepted in Linguistics, The Argument from Poverty of Stimulus(APS) is criticized for taken as granted and lack of convincing proof. 4 attempted proof on APS are provided in the paper and all of them are shown to be flawed. The authors never claims to advocate nor against APS, but stays neutral.
ReplyDeleteThe attempted proofs are attacked on their incorrect assumption, inaccurate data, etc. I would like to point out that, despite the attacks are on empirical analysis, the anti arguments themselves are empirical as well, as they are drew from texts sources of massive corpus and children books. Thus they might also lack the accuracy of what kind of language children would hear and acquire, i.e. the oral language used by their parents on daily basis, which is not written in those books.
Data-driven learning is proposed as the alternative way to APS. My interpretation on data-driven learning is learning by trails and errors(it is exactly how all other non-through-language learning works, learning through language only needs categories to be grounded).
The paper point out how irrational linguistics are on taking APS for granted.
Pullum and Scholz delve into the argument for poverty of stimulus, and provide counter-examples for common cases nativists claim support the poverty of stimulus argument. I appreciated the style of this paper. Many scientists rely on denying any strong points of theories they disagree with, because they feel that it somehow strengthens their own position. Pullum and Scholz readily admit that innate language capacities may exist, and that their paper is merely trying to show linguist nativists where their argument is weak. History has shown that often the truth lies somewhere in between opposing theories, and I appreciate that the authors do not try to polarize the issue unnecessarily.
ReplyDeleteMy current issue with both nativists and data-driven proponents is that they use small samples to support their research. When one considers the vast exposure a child has to language, it seems like a drop in the ocean to be looking only at small snippets of conversations between parent and child. Even methods that looks at groups of phrases in texts are really just a vague approximation of reality. It seems that with the technology available to us, we should be undertaking the daunting task of gathering a complete history of language exposure for a sample of children as they progress through language acquisition and development. While this task will no doubt be difficult and tedious, it will allow researchers and theorists to develop a more systematic, and realistic appreciation of the possibility for data-driven learning or innate learning. Hopefully, this data will allow theorists to move away from the current state of research, in which individuals find extremely specific (aka not very general) situations to “prove” their theories, only to have another researcher “disprove” the theory with an equally specific counter-example.
The possibilities that come to mind from the possibility for systematic data recalls the massive breakthroughs in genetics across species with the mapping of entire genomes, including the widely successful Human Genome Project.
Well, there goes the neighborhood. I haven't seen Chomsky look this foolish since he and Allan Dershowitz talked politics (did ya know they went to the same summer camp?)
ReplyDeleteAPS was supposed to nail the hammer in the behaviorist's coffin. Did Pullman and Sholz perform a seance?
S. H. told us that story about S.A.P. misunderstanding the poverty of the stimulus. Could Chomsky have misunderstood his own argument? Probably.
I feel like Pinker defines UG as those set of rules that just about every language community employs, like recombination, subject-verb agreement, discrete morphology. Chomsky's APS is a bit more nuance. In "Is the Man Who Is Tall Happy," Michel Gondry's homage to MIT's living piece of American history, you can find Chomsky explain the APS.
APS or no APS, or whatever APS might be, there's no denying it. We have a language instinct. A thirst to categorize, if you will. I wouldn't be surprised if linguistic stimuli promote neurotrophic factors, which would explain how the language can lateralize and re-lateralize. Somehow, language constrains our attention; something draws us to language as opposed to prime numbers or the ocean. That we thirst for language, even just out of the womb without having been conditioned, without the sort of positive reinforcement or modelling that would otherwise train our ear, proves humans are a linguistic species.
If I understand the argument from poverty of the stimulus correctly, it argues that there isn’t always enough evidence for a child to learn things about a language. So there are things about languages that require evidence to be learned. Now this argument is for a more innate way that language is learned. What puzzles me a little is that they do say that there are plenty of other arguments for linguistic nativism, but don’t really discuss them. I’m wondering then, why choose this argument? Is that because it (one of several) is the argument that seems the weakest because it does not have evidence to back it up. That I understand. What I did kept thinking about as I was reading was pidgin and creole. I’m quite interested in this topic because I find it fascinating that the children of different groups of people who only communicate with one another through a very simple language, can create a full grammatical language that often has features that are different from either of the languages of their parents. I do wonder what evidence then, leads these children to have a full grammar when all they have is the lexicon from the parent languages. I’m not sure if this fits into the argument from the poverty of the stimulus or maybe underdetermination where children develop grammars that are undetermined by the data, but I’m not entirely sure what they mean by underdetermination here.
ReplyDeleteEither way, perhaps this argument should be pursued in a more empirical fashion so that we can have some evidence to back up the claim.
To me, the greatest value of this line of argument is the clarification and restatement of APS provided at the start. After this, the paper seems to turn somewhat pedantic, arguing that children do in fact have some access to non-UG sentences, though the argument was never "no access" but the rather mushier and therefore harder to test "not enough".
ReplyDeleteIn the paper Empirical Assessment of Stimulus Poverty Arguments, Pullum looks through some arguments defending APS and find that they are not as tenable as they seem. However, he says he is not rejecting APS or endorsing empiricism, but only saying that if we want to attribute language to APS, we need to do more researches and find more evidences.
ReplyDeleteIn the arguments that Pullum refer to, most of them seem to say that we are able to learn some grammar although we do not have a full access to it, and Pullum says those arguments are not tenable. Despite whether what Pullum says is right or wrong or these arguments are right or wrong, I find that most of the examples seem to come from normal grammar, and I do agree that normal grammar is not innate but is learned after we are born (through mere exposure, trial-and-error and instruction). If normal grammar is not innate and is learned, it is very likely that those arguments cannot really hold if they only care about normal grammar because if they hold, APS holds, and language nativism holds. However, if we look at Universal Grammar, we will find that it is innate, but it is very hard to give out examples like in the arguments that Pullum refer to because everything we speak, should be right in Universal Grammar (as I mentioned in the first skywriting this week), and therefore we will not be able to give out examples that are not right in Universal Grammar (i.e. negative evidences) and I think this is the part Pullum fails to look at.
From recent two weeks' readings, I find that when we talk about language, no matter the evolution of language or language acquisition, Universal Grammar always seems to be so different that it makes many of our arguments incomprehensive. I guess in order to fully discover how language evolves and how we acquire language, it is essential to discover how Universal Grammar is evolved and how and why it is inborn in our heads. Also, to relate to the previous weeks' discussion, I think Universal Grammar might also play an important role in cognitive science - reverse-engineer about how and why we are able to do what we can do.
When Pullum and Scholz stated: "We ourselves would not charge someone with being out of their mind for at- tempting a full input corpus for a designated infant; it would be an extremely useful contribution to science. But Chomskyís last point is nonetheless correct: useful work can still be done by looking for cases where it is implausible to assume accessibility. The lack of a complete corpus of the input for any given learner will not be an insuperable obstacle provided only that corpora of En- glish text do not vary dramatically with respect to the constructions that they contain. Our default assumption will be this: a construction inaccessible to in- fants during the language acquisition process must be rare enough that it will be almost entirely absent from corpora of text quite generally", I found this to be a satisfying parallel with our class discussions of categorisation—more specifically, that cognition (and categorisation, as part and parcel thereof) is concerned only with categorisable kind of things. That is, if cognition begets categorisation, and categorisation cannot, by principle, deal with non-categorisable things, then cognition is unconcerned with those things as well. This similarity is obviously inconclusive and unproductive, in context of criticisms of the poverty of stimulus arguments, and the further resulting view that perhaps language acquisition is not so much a designed modular organ that would hold novel insight for the nature of cognition, but I do see the appeal of the latter, and therefore some for the former as well.
ReplyDeleteDistinguishing all UG vs non-UG utterances (and producing all and only UG utterances) is a categorization task. To learn a category, you need evidence of what's in it as well as what is not...
DeleteAfter today's class, I realized that the discussion about UG is far from resolved! Here are some big questions that remain, along with a few ideas that I am sure are wrong but would like to know why.
ReplyDeleteWhy are all languages bound by universal grammar? What is the adaptive mechanism? Without UG. is thinking impossible?
As mentioned, language allows us to express and share our thoughts. But if (as some might argue) without language, we cannot think, then does language enable us to think? Could in fact the adaptive mechanism for language then to be to allow us to think, and to all think in the same way? THen the question becomes, why must we all think in the same way? Well perhaps we do not think identically, but our neurological mechanisms that allow for thought are the same, and language allows us to ground what is going on inside of our heads. Another way of saying this that might make more sense is that perhaps language grounds the meaning of otherwise meaningless squiggles and squiggles (neurons firing) inside our heads.
Great apes can think too. But only our species can think verbally -- which I think means propositionally (true/false subject/predicate statements). So that's a special kind of thinking, unique to our species. It doesn't mean we all think the same thing, just that we can all think in the the same (propositional) way.
DeleteBut language certainly does not ground meaning. The meaning of the words has to be grounded to allow language: the words need to be T3-connected to their referents through our brains, otherwise they are just meaningless squiggles and squoggles.
But so far the connection between symbol grounding and UG, if any, is not obvious. Perhaps the NON-autonomy of syntax will turn out to have something to do with it...
“[…]there could be positive evidence for the correctness of some rule or principle, evidence that could be found in language use by an adult linguist doing a really intensive search, but would not emerge in conversational data anywhere near often enough to guarantee that any particular child would ever encounter it.”
ReplyDeleteThis article seemed to do a good job of clarifying the different ways “poverty of the stimulus” can be interpreted and argued. Previously, I had thought that POS necessarily referred to the extremely small amount of negative feedback received by acquirers of a new language, but now see that it can be interpreted to mean that there is simply not enough data with which the learner can form a specific rule.
If the framework outlined by Pullum and Scholtz for empirically testing APS is sound, I’m hopeful to hear about research utilizing their framework. If the data-driven approach turns out to be closer to what we do when we learn language, will it make building machines that implement our language learning capacity easier (i.e. might we need less modules)? Would the data-driven theory necessarily indicate that a language module is more integrated into general-purpose learning modules than if the innately primed hypothesis turns out to be true?
(Replace "language learning/acquisition" with "grammar learning/acquisition" where appropriate...)
DeleteAs many classmates have mentioned prior, Pullum and Scholz lay out 5 steps that must be satisfied in order for the APS to hold. As mentioned, in all 4 case studies provided by the authors, it is always the step requiring a lack of evidence that is brought into question, with the authors providing multiple counter examples showing that the inaccessibility of negative evidence is false. In every instance where theorists support the APS, their “rigorous” criteria merely demands that examples, showing a child what is structurally incorrect, do not occur “nearly enough” to influence a child’s language acquisition. The notion that examples of incorrect language are not provided “enough” seems quite ambiguous as true scientific evidence for UG. When Pullum goes on to show multiple examples of accessible evidence, it seems implied that experimenters cannot truly verify the notion of UG unless sample children are reared in environments where inputs are highly controlled and monitored by an experimenter, which of course is highly unethical. In my opinion, problems arise when trying to isolate innate behavior in a world that is fundamentally linked with data from the environment. The notion of UG is, of course, highly seductive. It explains why other species cannot possibly learn language to the same extent as a human. If language acquisition occurred based solely on the internalization of what we here, how would we have developed the capabilities that we have today. Apes communicating with apes with never result in an ape communicating the way a human does. Thus, internalizing what people communicate to us doesn’t leave any room for one to adapt and expand on the language! The existence of an innate UG would point towards genetic means that differentiate humans from other species. Unfortunately, as Pullum shows, verifying its existence requires more
ReplyDeleteI agree with some comments above regarding the difficulty of this article to get through. It's a little easier post-class discussion, but still will probably need to return to it several more times. In the meantime, the feature that stood out to me most was the "lacuna" referenced by Pullum and Scholz. They describe it as "the set of crucial sentences... such that if the learner had access to them, the claim of data-driven learning of the acquirendum would be supported." There are a lot of interesting things to be said about how lacunas relate to the linguistic arguments and theories put forth by the authors and other researchers mentioned in the article, but all I could think about was how analogous the lacuna is to Professor Harnad's kernel - it seems to me that they are almost the same thing; though the kernel has to do with symbol-grounding and the lacuna has to do with language-learning. Furthermore the lacuna includes sentences while the kernel is meant for words. Still, I think there is potentially room for many interesting claims to be made regarding the similarity of these two concepts; maybe knowing more about one can inform further studies into the other.
ReplyDelete“On the other hand, we have noticed that some features of English have the opposite property, being much more frequent in spoken language than in printed prose ….. They are extremely rare in print …. For a construction like this, it would be easy to under estimate the frequency of examples in the spoken language if one looked only at written sources…. We assume that research on one kind of text will tell us at least something about the syntactic structures likely to be found in others. This seems to us a sensible assumption because the main determinant of how often particular constructions turn up will be the random exigencies of what needs to be said, and we take it that these random exigencies will be roughly constant across many genres.” (22-23)
ReplyDeletePullum and Scholz reading calls for a much needed accumulation of evidence that there exists in fact the APS. As they review repeatedly in their paper, some utterances that linguistic nativists would think of being extremely rare for children to be able to know they are grammatically correct, turn out to actually appear in the data banks. What I found most interesting is the issue of where to set the threshold, what amount of information can account as enough evidence for language development to be actually data-driven and not UG? Linguistic nativist make the case that the information that is already available to the child is not enough, and furthermore that there exists an absence of negative evidence (argument that Pullum and Scholz intentionally avoid).
Harnad writes in the commentaries for Pinker’s reading: "Maybe it requires less grammatical accuracy to understand than to speak". The possibility that the grammar necessary to understand is less strict than that one used for language is a very interesting idea to me. Particularly here in Montreal, we can observe this everyday, you find people with first languages that are different from English, in every conversation there might be some English grammatical errors, but these kinds of mistakes do not interfere with the conversation, because whatever needs to be understood is understood. (In my case I would not stop somebody from telling me something just because they made a grammatical error, as long as I understand what they are telling me). If we can understand without perfect grammar, could this account for a “meaning organ,” rather than a “syntax organ”? To connect this with what I said previously, could we argue that there is enough evidence about meaning in both the sensorimotor experiences and the propositions and utterances that happen in parallel with those experiences around children? (I know I might be stretching the string here a bit).
Andrea, I doubt there is both a language production "organ" and a language comprehension "organ." There's only one for both. But it is apparently possible to learn enough passively, without speaking, to understand, up to a certain point. And for this passive, approximate comprehension, perhaps grammar errors matter less. In any case, I suspect that most of the grammatical errors children hear are ordinary grammar errors, not UG errors. It is unlikely that children who do not yet speak have enough understanding to be able to distinguish UG-compliant from non-UG-compliant utterances. But who knows? Maybe they can already tell them apart. In any case, once they do start speaking, they don't make UG errors.
DeleteIn relating language acquisition (and UG) to cognition, I now understand that it has to do with generalization and lack of negative evidence. The role of UG (constraining all human conventional grammars) is the same as our cognizing ability. That is, our capacity to make linguistic generalizations solely from positive evidence relates to how we cognize about other things in the real world.
ReplyDeleteHowever, I think this generalization takes it one step too far (i.e. Generalizing UG to a general capacity of cognition rather than a capacity that is specific to language). To provide support for the above claim, language deficits such as SLI affect one’s linguistic abilities and not their general cognitive abilities. This shows that although language is connected to cognition, it is completely separate from it in that it has its own distinct module. One could argue that these deficits leave UG intact and affect other structures that influence our ability to acquire language. If this is true, then I may have answered my own question.
I found this article to be quite a challenge to read and understand, given that I don't have a background in linguistics. However, I think I did understand the general ideas it put forward.
ReplyDeleteMy eventual conclusion is that I think it's a bit ridiculous that no one has followed an infant from birth until fluency yet. It's really not that hard a task. Mind-numbingly boring, perhaps, but not really. There are plenty of people in this world who make careers out of transcribing audio who I'm sure would appreciate the job opportunity. Moreover, with speech recognition technology these days, the workers would only have to fix errors in generated transcriptions. In my opinion, recording and transcribing a child's interactions for the first few years is a very small price to pay for advancing science and finally showing once and for all whether or not UG is a sound theory.