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.
Only for the stout-hearted:
Everaert, M. B., Huybregts, M. A., Chomsky, N., Berwick, R. C., & Bolhuis, J. J. (2015). Structures, not strings: linguistics as part of the cognitive sciences. Trends in Cognitive Sciences, 19(12), 729-743.
(And a critique, but is it valid?):
Dąbrowska, E. (2015). What exactly is Universal Grammar, and has anyone seen it? Frontiers in Psychology, 6, 852.
Hi! The link above does not seem to be working for me. I found the same article on this website that is accessible through logging in to the McGill VPN: https://www.degruyter.com/document/doi/10.1515/tlir.19.1-2.9/html
ReplyDeleteKimberley, thanks, I've updated the link. (There are over many OA versions in Google Scholar.)
DeleteIf what we've been discussing in class about categorization and learning is true, then it seems that the argument against the poverty of stimulus (POS) put forth in this paper in entirely misconceived.
ReplyDeleteThe reason why UG must be inborn is that children never produce or hear UG non-compliant sentences, which in turn makes supervised learning impossible. To go back to the island example, in order to learn what mushrooms are edible/not-edible, we need to be exposed to both edible and non-edible mushrooms. If we were only exposed to the former, we could not identify the features that distinguish the two types of mushrooms. Likewise, without being exposed to UG non-compliant sentences, we could not learn what distinguished them from UG compliant-sentences.
So POS is all about the absence of negative evidence, not (as the authors suppose) the scarcity of certain types of positive evidence. Their argument turns out to be irrelevant.
Gabriel, that's right, and it's the reason I assigned this reading. It illustrates the kinds of misunderstandings there are about UG, even among linguists.
DeleteThis paper was complicated and did not clarify much as to why cognitive science needs to understand auxiliaries and structure-dependence vs structure-independence. I had never thought before about cases in which the first part of the compound is itself complex, like “new books shelf,” in which one comes to understand it as a shelf of new books rather than a new shelf of books. The paper claims that our experience is not what allows us to discern what is prioritary in our sentences since we don’t receive explicit negative evidence. It also emphasizes the influence of our environment which can be highly variable, but yet we all still have inherent domain-specific knowledge that stays fixed whether we like in rural villages or modern, developed cities. Will there ever be an answer to how typical learners know to avoid UG mistakes?
DeleteTess, like P & B, Pullum is fuzzy about UG, by not distinguishing UG rules (unlearnable by child because no negative examples – not “too few”: none, for the child ) and OG rules (plenty of positive and negative examples, learnable by unsup and sup learning as well as verbal instruction). The “big data” survey of texts is irrelevant (and would still be irrelevant today, in the GPT-3 era.
DeleteI found it interesting that different syntactic structures are used in some forms of communication and not in others. The author illustrated this with the example that consecutive synthetic and analytic negotiations, such as “We can’t not go”, are rarely found in printed sources, but they are quite common in conversation and film dialogue. It’s also demonstrated that this genre difference cuts both ways, as Biber found that about 31 past participial reduced relatives are found per 10,000 words in journalists’ prose, but only one occurs in 10,000 words of face-to-face conversations. Why is there such a discrepancy between the way we spontaneously communicate and how we write? We certainly learned the conventions of spoken language earlier on, as we learn to speak before we learn to read or write, and I wonder if the emerging ability to read and write changes something about the way we communicate naturally. Is this difference just due to the more formal conventions of writing or is there something else?
ReplyDeleteAlexander, yes, spoken use has primacy over written use, but neither is relevant to the child’s short language-learning input and output.
DeleteThis article aims to refute the validity of the poverty of stimulus argument as defined by Chomsky and other theories of cognitive science as a component of the nativist view of language.
ReplyDeleteI found their argumentation quite unusual, as they refute the argument by redefining it on their own terms. The poverty of stimulus argument (PSA) they use is based on the constraint put on language-learning for infants by the lack of crucial evidence provided in a data-driven acquisition of language (unsupervised/supervised learning). By doing this, the authors avoid the issue of language acquisition crucial to the PSA, which is the lack of ‘negative evidence’/UG non-compliant sentences (as explained in Gabriel’s comment) in infants’ environments. The authors argue that this issue is not within the scope of their article, and that they do not deal with it for lack of space. They also suggest renaming the argument as the “Argument selected by Pullum and Scholz” to emphasize that they are dealing with a revision of the PSA, which I found bizarre as well if their aim was to refute the ‘original’ PSA.
The authors then apply the PSA to various language phenomena (such as regular plurals in compounds, auxiliary sequences, the anaphoric one, etc), aiming to advance a theory that is supported by more empirical evidence than Chomsky’s PSA. I wonder about the validity of this argument in the paper: how much empirical evidence is provided in support of the PSA, and does it need more to be considered valid? More generally, what kind of empirical evidence can be provided for the innateness of a cognitive capacity?
Mathilda, Week 9a and 9b readings are meant to illustrate how even well-informed authors keep getting it wrong, about UG and the PoS. I hope that at least the 48 students in Psyc 538 will get it right, and understand exactly why and how these authors (and others) keep getting it wrong.
Delete(The surprising thing is that it is all relatively simple! Understand that there is a UG/OG distinction, and that what distinguishes them is PoS: The child neither produces not hears UG violations or corrections, so UG cannot be learned. The important thing is to recognise that you don’t really know what the unlearnable UG rules are [if you haven’t studied syntax], so don’t substitute fantasy-rules for UG rules when you make the UG/OG distinction, because your UG rules will inevitably be OG rules, hence wrong.)
The main takeaway I got from this paper is the argument of the poverty of the stimulus (POS). POS is the idea that language output during language acquisition is underdetermined by the primary linguistic data that is available in the environment. This is the basis of the POS argument. It’s the fact that only UG-compliant sentences are heard and spoken (even during language acquisition) that points to the underdetermination. The fact that “some sentences never occur” just consequently shows how UG is underdetermined and cannot be learned. As we saw from our categorization lectures, there are three ways of learning: unsupervised, supervised and hearsay/instruction.
DeleteThe absence of negative evidence makes unsupervised learning impossible because everyone speaks UG-compliantly. Children cannot passively induce the features distinguishing OG when all they receive (are exposed to) is the features distinguishing OG and never features that does not distinguish OG.
The absence of negative evidence makes supervised learning impossible because there can be no error correction if children only speak UG-compliantly. Thus, because this “absence of negative evidence” (poverty of the stimulus) shows UG cannot be learned, it leads to the logic that such grammatical rules must be inborn.
As Gabriel mentions above, this paper seems to have missed the idea behind Chomsky's POS argument as we studied it in class. Indeed, they state that they are "setting aside arguments based merely on cases in which it appears that children would need negative data", which corresponds to the correction of ungrammatical sentences. Setting this aside is inappropriate since the POS argument that we mentioned in past weeks refers to the lack of these 'non-members' of grammatical language (ie examples of ungrammatical sentences). As such, the authors are wrong to focus on children's lack/insufficiency of positive evidence examples (ie members of grammatical language, aka grammatical sentences).
ReplyDeleteAmélie, the child’s lack is the lack of UG violations (the lack of negative examples of UG, whether produced by the child or by adults, hence also the lack corrections). There is nothing to correct except OG errors, and even much of OG can be learned by unsupervised (observation/imitation) learning alone, without correction.
DeleteCan OG be learned through positive evidence only? In the 9a skywriting, I convinced myself that negative evidence was necessary for learning.
DeleteIf everything is an example of positive evidence, it would be unclear what said evidence is actually evidence of. In other words, a category isn't really a category if everything fits into said category. Its features are no longer distinguishable from non-members, if everything is a member, and feature detection is necessary for category learning.
But if a portion of OG can be learned unsupervised, without correction, how do we know what's OG and what's UG? Or is this response to Amelie's post just to say that sometimes some children learn some rules unsupervised, but overall, all OG rules can be learned with supervised learning.
Teegan, yes (to your last sentence). A lot of OG is simple enough so unsup learning and imitation is enough.
DeleteHi Teegan, I am not sure, but from my understanding, OG could be learned from positive experience only (even if it is not the case). Finker's learnability theory applied to language suggests that children form a hypothesis on correct language use and tend to have an overly large hypothesis (for example, one that contains both went and goed). The child can narrow down its vocabulary by positive evidence alone, that is by listening to their parents' correct utterances and correcting their hypothesis. As per the distinction between OG learned by positive evidence alone and UG, I believe you're right in saying that OG can also be learned with supervised learning. Its rules have been studied and can be explained fairly simply, unlike UG which cannot be learned.
DeleteRosalie: Yes, some OG rules can be learned from positive evidence and unsupervised learning. This has no bearing on UG.
DeleteSome OG needs supervised learning, and that needs negative examples and correction. But that too has no bearing on UG, where there are no negative examples.
OG can be picked up passively, but not UG because of poverty of stimulus, no matter how much exposure you get. But they can be learned via positive feedback, i.e., supervised learning. But how come we sometimes will make some expression errors, but not often; do we consider that to be a general OG error or unsolid UG capacity?
ReplyDeleteIf UG does not change, then this error would be consistent throughout the person's life, so it can't be UG error?
DeleteMonica, the child's errors are all OG errors. There is no error-correction for UG because the child makes no UG errors.
DeleteI found this paper quite difficult to grasp however, understood it to be an explanation to why language is likely to be acquired through empirical means (data-driven) approach as opposed to the nativist belief support by the Argument from the Poverty of the stimulus (APS). The APS states that human infants learn first languages by means of innately-primed learning, as they would lack all the experience necessary to completely learn language. Data-driven learning, on the other hand, completely relies on generalizations made from experiences by ordinary methods. What (I think) was alluded to in this reading, was the APS argument can be weakened by demonstrating data-driven learning/success through experimentation. Overall, I am left questioning if this paper is trying to prove that there is no validity to APS (by saying it is difficult to defend). My belief is that the truth lies in the middle (based on our past readings and discussions on topics such as supervised and unsupervised learning). As children, we definitely are not experiencing every stimulus to fully form language, but I believe that we build upon our innate ability through data-driven learning.
ReplyDeleteIf I am understanding correctly, the conceptualization of universal grammar is that is inborn, and thus it would never be possible to make an error or be corrected. No learning was required in order for the individual to have perfect UG. We never hear any UG errors because every human was born with the ability to perfectly use UG.
DeleteKimberly, actually we do hear UG errors, but never from the language-learning child (except if they imitate from eavesdropping on their parents, if they are both linguists, talking shop at home).
DeleteThe way linguists piece together the features of UG is by testing their hunches on what might be a UG rule by trying it out with examples they produce: “If this is a rule of UG, then I ought to be able to say this, but not this.” If the hunch is right, then deliberately saying something that violates the rule should be wrong. How do they determine whether it’s wrong? Because it sounds wrong (“John is easy to please Mary”), just the way violating OG sounds wrong (“He don’t talk English right”), except that with the OG violation the rule being violated is one that we learned (through unsup or sup learning and corrective feedback, or, by verbal instruction from (ordinary) grammar schoolteachers), whereas the UG violation is one that our ear/brain detects innately when we test out a hunch about what might be a rule of UG.
If you think about it, this is a rather remarkable way of doing experimental science. The way that usually goes is that you have a hypothesis, and then you test if it’s right by doing an experiment to see whether it confirms the predictions of the hypothesis.
But when linguists test a hypothesized UG rule, they just have to consult their own ear/brain to see (or rather hear) the outcome! This is because having an innate UG in your brain means you can tell whether a sentence complies with it or violates UG, just as you can with sentences complying with or violating a learned OG rule.
CHALLENGE: I haven’t yet thought about this myself, yet, but can any of you think of other examples of doing experimental science in this way – by doing mental “thought experiments”?
Searle’s CRA thought-experiment sounds somewhat similar “If there were a computer executing a T2-passing computer program that could supposedly be understand Chinese, would I be understanding Chinese if I were executing that computer program?” But the CRA’s outcome is deduced by logical reasoning, which makes it more like mathematics or logic rather than experimental science. With linguists doing hypothesis-testing of candidate UG rules, however, there really does seem to be empirical uncertainty about the outcome until you actually do the experiment, produce the sentence, and then listen to ”hear” whether the outcome sounds right or wrong. Empirical science is really “observing” the outcome data, to resolve the uncertainty.
Can you think of other examples?
Although this method of defining UG rules sounds remarkable, how do the linguists who are hypothesizing the rules and subjectively rating the UG-compliance of sentences they utter remove their personal linguistic biases? Does the empiricism of these experiments rest on the assumption that we all have the exact same innate UG category?
DeleteHi professor, another thought experiment I could think of is Galileo’s gravity experiment with the Leaning Tower of Pisa. Although it is often claimed to be a real experiment, but it simply a hypothetical proposed by Galileo to explain how gravity works on objects of different weights. If dropped a heavy ball and a light ball off the Leaning Tower of Pisa, which one would land first. Others think it is the heavy, but Galileo disagreed. He said that both balls would hit the ground at the same time because the degree of acceleration is unrelated to mass. Other than this one, I’m not sure if Mary the scientist could also be another example.
DeletePolly, linguists check their grammaticality judgments with other linguists (and nonlinguists). And they publish them. Agreement is very high…
DeleteNadila, I think Galileo’s was a real experiment, not a thought experiment.
To answer your question professor, I believe another example of conducting experimental science this way would be in identifying dissonant versus consonant chords. We can naturally identify dissonant notes (and usually react badly) when we hear them as opposed to consonant chords. Although notes can be broken down mathematically, you maintain this capacity even with no knowledge of mathematics. Whether or not this is a learned capacity (even through unsupervised learning) or if this capacity is innate is up for question. I am leaning towards the latter just because I feeeel like if you played a dissonant note to a child, they would react unfavorably as opposed to consonant chords. I also think it goes back to the idea that we can somehow sense the "emotion" of the an instrumental melody even when nothing is explicitly said and this quality is universal over different cultures. What I find interesting is that both of these mental "thought experiments" (UG error identification and note identification) rely on our auditory sense. I could not really come up with one that relied on another one of our senses so maybe I'm just not creative enough or perhaps there is something special about our auditory capacity.
DeletePullum & Scholz dispute the argument from poverty of the stimulus (APS), arguing it to be insufficient in validating the claims given by linguistic nativism. Their argument, however, is based on their fundamental misunderstanding of what poverty of the stimulus is. This appears to stem from proponents of linguistic nativism who also misunderstood the APS and failed to differentiate between OG and UG. P&S believe that the APS is about positive evidence not emerging in conversation enough for any given child to encounter it. They fail to address negative evidence, which is the very root of the APS. They argue that children receive more input than previously thought, not just from language directed at them but also from indirect sources. This argument only addresses OG, and completely misses the point of the APS. OG is learnable, while UG is not. OG has positive and negative examples so children can get corrective feedback, whereas UG has a poverty of stimulus: there are no UG-violating examples so children cannot learn it (it is innate). Similar to Pinker, P&S conflate OG and UG, which makes their argument against linguistic nativism useless. It seems like multiple authors in the field, whether opponents or proponents of linguistic nativism, fail to understand the very thing that they are arguing for/against.
ReplyDeleteJosie, that’s right: Now explain to all the misunderstanders why both positive and negative examples (members and non-members) are needed to learn a category (whether by unsupervised or supervised learning). -- And explain also how this is just as true for “mirrored” perception/production categories (ba/da/ga, smile/frown, chess-moves) as it is for purely perceptual categories (apples, mushrooms, striped).
DeleteIs it correct in saying that UG is thus proved through negative evidence whereas OG is determined through positive and negative examples?
DeleteEmma, I’m not sure what you mean by “proved,” but UG is inferred from evidence about what children can do: Since they can speak UG-compliantly without errors or corrections they must be born with that capacity. In contrast, they can and do learn OG, by unsup and sup learning (mistakes and corrections) and verbal instruction.
Delete(“Negative evidence” and “negative examples” are used interchangeably to refer to that children hear and say, and any corrective feedback they get. To learn a new category you need examples of what’s a member and what’s not, so you can learn what features distinguish them.)
Both positive and negative examples are needed to learn a category (i.e. distinguish between members of different categories). Through unsupervised or supervised learning, exposure to positive and negative examples makes between-category differences look bigger than within-category differences. For example, when learning to categorize edible mushrooms, you need to have examples of both edible and inedible ones before you can form the category. If you were only provided with edible mushrooms (positive examples), then you may come across an inedible mushroom (negative example) and eat it because you cannot distinguish it from edible ones. Post+neg examples are just as necessary for learning “mirrored” perception/production categories (smile/frown) as for learning purely perceptual categories (e.g. mushrooms). While perception/production categories have features and affordances that are sensorimotor (and can influence each other), these features/affordances still distinguish one category from another and require both negative and positive examples to do so. When learning to categorize a smile vs a frown, for example, we may perceive and “mirror” positive examples of a smile, but still require negative examples to know the category and be able to distinguish a smile from a frown. This demonstrates to misunderstanders that negative examples are equally as important as positive examples in learning a category.
DeleteJosie, that’s right. You cannot learn the features that distinguish rabbits from hares from observing only rabbits or only hares. Same for learning the difference between ba’s and da’s, both their perception side and their production side. And the reason categories look more different after you have learned to detect their distinguishing features is that you have learned to ignore their non-distinguishing features.
DeleteEven though I'm less bullish on Universal Grammar than many, I agree with many of my classmates in that this argument doesn't engage with Universal Grammarians in a fair or productive way.
ReplyDeleteThe theory of Universal Grammar doesn't claim that children never learn by external evidence- that would be foolish. We obviously learn our languages via our community and especially our parents. That is 'OG.' We get make mistakes and get corrections when we mess up surface grammar conventions.
UG is literally defined as those grammatical constructs that are not learned through observation; the errors that children do not make, ever. There is no evidence from childhood learning suggested by Pullam & Scholz that suggests that any of the invariable constructs put forth by UG proponents is wrong. They are just showing that OG is reinforced, which nobody disputes.
If we do want to critique the poverty of the stimulus, we can. It could be possible to show (empirically!) that there is some inbuilt structure when using a language that is not necessarily contained in the mind, but one has to meet Universal Grammar on its own terms to do so.
Jacob, Neither K-S nor I understand your last paragraph.
DeleteI was attempting to make the point that poverty of stimulus is debatable. However, that doesn't necessitate that UG is wrong entirely- there may be necessary constraints on the way language works and an inbuilt structure whether children have PoS or not!
DeleteJacob, K-S does not know what is debatable about POS: (1) no UG errors? (2) no agreement on UG errors?
DeleteWhat does "necessary constraints" mean?
What kind of "inbuilt structure" (if no POS)?
It's intriguing to me that many linguistic professionals negate the distinction of UG in their articles. It posited the question for me of what would change about the way we interact with language (teach it, speak it etc.,) if we understood it?
ReplyDeleteEmma, being taught the rules of UG by linguists does not seem to be necessary for learning language! But in cogsci they are important for understanding how to reverse-engineer it.
DeleteThis may be a dumb question but do you think if more linguists (not just the ones at MIT) actually understood the syntactic rules of UG they would understand the POS argument better? I know this shouldn't matter but maybe the 'technicality' of it would allow them to more systematically see the lack of negative evidence that children receive when learning UG.
DeleteSophie, I think the answer to your question is: Yes. But I also think that it is possible for the errors -- of the linguists (and nonlinguists) who do not understand UG well enough -- to be understood by linguists who also do not yet understand UG well enough (and even by nonlinguists [like lilliputian me] who do not know UG at all). (Think about it: Unless I have been teaching you nonsense about the difference between OG and UG [which is always a possibility!], those who have understood this course will have understood the errors --- without knowing UG!)
DeleteThe article written by Pullum and Scholz breaks down the argument of universal grammar, specifically the existence of empirical proof for the absence of the stimulus i.e. proof that we have innate grammar abilities and are able to learn rule that are not explicitly stated. The authors essentially refute major empirical supports for innate grammar knowledge by providing in-text examples that counter any suggested absence of evidence to learn a specific grammatical rule. The article was essentially a long repetition of counter examples, that lead to the conclusion that there is no evidence for the absence of the stimulus, innate grammar knowledge, and thus universal grammar though they warn that they are not arguing that language is definitely data driven, rather than innate.
ReplyDeleteLaura, please read the replies to other comments on 8a, 8b, 9a and 9b. What did Pullum leave out?
DeleteI am not sure what else to add to this thread, since a lot has been said. It is clear that, similarly to week 8a (Pinker) and 9a, this article is again just a misinterpretation of how grammar evolved… Or at least another example of how people seem to ’forget’, ignore or misinterpret universal grammar. More specifically, this article aims to refute the poverty of stimulus argument but misses the central idea to such, which is that there does NOT exist ‘non-members’ of the UG category. In other words, it is not the case that people make sentences that do not comply with universal grammar rules. From my understanding, this is the central point that Pullum left out.
DeleteVitoria, not quite that simple. Please read the replies about linguists deliberately producing UG-violating sentences in the process of trying to infer what the rules of UG are.
DeleteAs many have summarized, this article is missing the point of not differentiating between UG and OG. They raised many examples to refute arguments supporting POS. For instance, on page 26, they argued that British English's advent use of plural noun-noun compounds ("a drugs problem") is incompatible with "the view that universal grammar dictates the principle involved." To my understanding, this merely shows the diverse dialect variants of OG for the same language.
ReplyDeleteJenny, yes, dialect (OG) differences, OG changes, and idioms.
DeleteCynthia, you are right to ask how we recognize the difference between UG+ and UG- utterances. The answer is crucial for how linguists study UG. They must make “grammaticality judgments.”
ReplyDeleteStart with OG: Native speakers can tell if they have heard an OG- error (e.g., “between you and I”) even if they don’t know the rule explicitly, because it “sounds wrong” to their learned implicit OG feature-detectors. They learned to detect the difference through unsupervised and supervised learning, both for perception and for production. (Both UG and OG are mirror capacities.)
It's the same for UG, but without the learning. Everyone has implicit UG+/UG- feature-detectors innately. Linguists call these bases for their grammaticality judgments “syntactic intuitions.” If they did not have those intuitions for testing whether or not a rule is part of UG, they could never figure out what the rules of UG are. But if they could not tell the difference between UG+ and UG- then there would be no UG at all! The same production filter that produces only UG+ output can distinguish UG+ input from UG- input.
This is not to say that linguists don’t encounter fuzzy cases, and even cases on which they disagree. But these cases are few; linguists’ (and nonlinguists’) grammaticality intuitions largely agree. And some of the UG disagreement is no doubt because of OG differences and first/second language differences rather than UG.
(Some linguists have suggested that grammaticality judgements are not +/- categorical judgments but continuous 0% to 100% degree-of-grammaticality judgments. If that were true perhaps we should expect some production errors for the weakest cases, rather than just judgment errors. But the disagreements are so few that it does not seem to make much difference to most of the project of making the rules of UG explicit.)
After reading this paper entitled “Empirical assessment of stimulus poverty arguments” by Pullum and the reading in week 9a, I find it amazing how there are misunderstandings about the poverty of stimulus argument.
ReplyDeleteWe defined the “poverty of stimulus argument” (POS) as how children are never exposed to UG violations. This means that children are never exposed to sentences not obligating the rules of UG; hence, they are not disclosed to corrections. You need correction to learn, as learning categories is done by receiving feedback on a trial-and-error basis. What the authors have done differently here is that they have tried to change the POS by claiming that it is not about the lack of violations and negative examples of UG but rather that “the relevant positive evidence, although it does exist, is not accessible to learners during the acquisition process, because of its rarity: linguists can in principle discover it, but children will not.”
This differs from the view we have seen in class, and I think this doesn’t seem right as children are exposed to so many different stimuli and sources of input that they must encounter relevant positive evidence. The problem is not with the rarity of positive inputs but rather the lack of negative inputs for UG with children.
Étienne, you're almost right, except that far more important than the fact that the child does not receive UG- input, only UG+ input, is the fact that the child never produces UG- output; it produces only UG+ output. Hence there are no (UG) errors to be corrected (only OG errors). And that's the reason we have to conclude that UG is innate!
DeleteAs many have mentioned above, this text along with Pinker (9. a), are examples of arguments that fall short when not correctly understanding the distinction between OG and UG.
ReplyDeleteThe fundamental distinction between these two forms of grammar is that one is learned through childhood (OG), and the other is innate due to the poverty of stimulus (UG). OG rules are learned by supervised, unsupervised, and verbal learning. For UG, the rules are unlearnable; there are no negative examples (there are no examples that violate UG just floating out in the world).
This is why when it comes to language acquisition and evolution, it is easier to come up with theories for OG, but when dealing with UG, it ends up being a lot harder of a problem (which authors such as Pullum seem to skip by in the attempt to disprove any sort off nativism).
Thank you for the clarification! What I meant to communicate was there are no negative examples of UG- heard on a regular basis, kids do not make these mistakes. However, there are linguists that have come up with some examples: One you mentioned in class is:
DeleteJohn is eager to please Mary (UG+)
*John is easy to please Mary (OG-)
Prof Harnad, could you please remind me what the second example you mentioned is?
Kayla, "*who did he think that went out?"
DeleteIt works in Hungarian -- Kiröl gondolta hogy ki ment? -- because "who" can be inflected with a postposition to mean "About whom,” as in the Saxon vestiges of Elizabethan English: “Whereof did he think that (he) went out?.” But the Hungarian does not need the “he” because it has the pro-drop (pronoun drop) parameter (though “ment” can mean either “he went” or “she went” because Hungarian has no gender – although the culture is extremely sexist (so I doubt pronouns make any difference…)
While the authors are refuting the poverty of stimulus idea in this reading, I find the reasoning behind it quite irrational. There is a major focus on the property of positivity, where children are only exposed to correct UG, and never to ungrammatical UG. However, the main idea behind the PSA is that children are also not exposed to all grammatical UG (there is a poverty of the grammatical UG stimulus), and somehow, they still don’t make mistakes. The authors’ argument can be refuted back as: although the child hears neither the UG-compliant sentence: “John is easy to please” nor the UG-noncompliant sentence “ John is easy to please Mary”, the child still does not produce the non-grammatical sentence, why? Therefore, I believe either the authors are approaching this argument from an incorrect perspective, or they completely misunderstood Chomsky’s argument in the first place.
ReplyDeleteAlara, "Stevan Says" you are spot-on..
DeleteIf nothing else it is a little comical that in a 40 page paper aiming to refute the poverty of the stimulus (POS) argument, UG is only mentioned twice. As my peers above have pointed out, there is a gross misunderstanding in the authors conception of the argument, where they frequently conflate UG and OG. As Alexander above said, they argue that POS is caused by the lack of positive evidence, which was never the problem, the problem comes with the lack of negative evidence for UG non-compliant sentences. They also fail to distinguish that OG is not innate, and Chomsky never argued that it was. It is also a bit ironic that in the first page they critique many authors' understanding of the POS argument, only to not even understand it themselves!
ReplyDeleteI think I got a bit lost in the middle of their paper as well, the technical aspects seemed a bit redundant, so if there is something else I am missing from this paper please let me know
Sophie, you're not missing anything!
DeleteI found this reading interesting as it allowed me to reflect on computational models of (OG) language learning that I learned about in a computational psychology class last year. The models that I learned about (MINERVA, BEAGLE, LSA, among others…) agreed with this article’s misinterpretation of the PoS. P&S (and the theories underlying the computational models I looked at) understand the PoS as concerning how children do not have enough positive language exposure to explain their linguistic abilities to acquire OG-compliant language rules in a short amount of time.They thus refute the PoS on the basis that children instead rely on contextual information and generalizations made from repeated exposure to words/phrases/sentences (i.e., data-driven approach). However, the thing that these models miss is the fact that PoS cannot be refuted by modeling the association of word exposures across experience, as this only accounts for positive instances of both OG and UG. Only OG is learnable, as explained by Josie above, as children are exposed to both positive and negative examples. This is different to UG, which lacks negative examples found through experience, and thus is unlearnable (learnable only by linguists, as we learned in lecture). The lack of negative experience is crucial to understanding how UG is innate, and thus must be accounted for in these models if they seek to properly model human language capacities.
ReplyDeleteDarcy, all correct; but what's most important is the lack of UG- output from the child (speaking) rather than the lack of UG- input to the child (hearing).
Delete(Missing also is the corrective feedback that there woud have been from other people if the child had produced UG- output, but they don't. In contrast, they do produce OG- output, and they do get corrective feedback, though a lot of OG can be learned from unsup learning alone, without need for the sup learning feedback; this has often been another source of misunderstanding, along with conflating OG and UG.)
Pullum and Sholz challenge the poverty of the stimulus (PoS) argument by using a series of counterexamples to demonstrate that UG is not innate. However, the authors fail to differentiate between OG (learnable grammar rules, obtained through verbal hearsay, supervised learning, and unsupervised learning) and UG (unlearnable grammar rules). The latter are unlearnable because children never make UG errors nor hear UG corrections, however, they flawlessly produce UG-compliant language. Can we call UG a category if the majority of us have no knowledge of its non-members?
ReplyDeletePolly, UG+/UG- is a category, but it's a mirrored (perception/production) category (like chess).
DeleteIf you know the rules of chess, you don't make only C+ moves but not C- moves. And you can also perceive if someone else makes a C- move. The big difference is that C, like OG is learned and learnable (all three ways: unsup, sup and instructions) whereas the child knows UG without any of the three: it need only learn some OG.
Your chess analogy makes the distinction much clearer for me, thank you!
DeletePolly, actually I didn't fully answer your question! Yes, UG+/UG- is a category, and the reason the children don't get or need UG- examples to learn the rules (features) that distinguish UG+ from UG- is that UG rules are unlearnable without UG-, hence they must be innate.
DeleteIn contrast, the rules that distinguish OG+ from OG-, or C+ from C-, are learnable and learned, by unsup, sup or verbal instruction, all of which depend on sampling both + and – examples, which the learners do get.
The conflation between UG and OG is starting to become irritating. Although disputing the Poverty of Stimulus argument with negative evidence is a clever way. There hasn't been any evidence for there being any negative evidence so far.
ReplyDeleteThis leads to my main question. Albeit children do not make any UG mistakes, what about infants? Sure, we may not have any ways to measure it since they do not use any words, but I wonder if there was a way to find out if, in their babbling, they were showing some sort of structure relating to the presence of UG.
Alexei, see the other comments and replies. There is no UG- production by either the child or the adult, only OG+. So the child never hears UG- either, so never gets corrective feedback. As it learns OG, it makes OG- errors, and gets OG corrections and even instruction (though it can learn a lot of OG with just unsupervised learning too).
DeleteI don't know the answer to the question about babbling, or even the first primitive steps at speaking -- which are complicated by the fact that the child understands somewhat before it speaks, though it's hard to judge how much of that is a clever-Hans effect, or even a "clever-Bunny" effect ( https://www.youtube.com/watch?v=z8k2upr9vCE ), for Bunny fans, including me.
(Bunny's brilliant [so are horses], but she neither speaks (or, rather, stomps) nor understands language. "Stevan Says" this is because she cannot adopt (or has no interest in adopting) the narrative "propositional attitude" of subject/predication, True/False). That is not to say that she does not or cannot purposively convey information; just that she can't -- or is not motivated to -- transmit it propositionally. When she is told something propositionally, she can easily extract enough of the content associatively.)
This reading refutes the PoS (Poverty of Stimulus) argument. It presents four different counterarguments that serves as a basis for the rejection of the poverty of stimulus argument. However, there seems to be a gap between the poverty of stimulus argument as we’ve studied in so far and the way it is presented in this article. Indeed, Pullum & Sholz define the poverty of stimulus argument as the fact that “the relevant positive evidence, although it does exist, is not accessible to learners during the acquisition process, because of its rarity: linguists can in principle discover it, but children will not.”. In class, we have seen that this argument is essentially about how children lack exposure to UG negative evidence. Thus, the article doesn’t acknowledge the distinction between OG and UG. Their argument pertains to explain OG acquisition in children, but it falls short to explain UG as there are no negative examples of UG. More importantly, this reading doesn’t address why does the child not produce an UG- sentence.
ReplyDeleteInes, in trying to infer the rules of UG, linguists produce negative examples (UG-). How? But children neither produce nor hear them.
DeleteThis is a question I’ve been struggling to answer. How is it that linguists can produce these negative examples? The conclusion I’ve come to is that linguists can tell that their examples are wrong because they SOUND wrong. Like Prof. Harnad mentioned, linguists are trying to infer the rules of UG… the process they use is as simple is this: say they have a hypothesis for a rule of UG… if they want to prove that their hypothesis is TRUE and is in fact a rule of UG, then they must come up with an example that deliberately violates that rule. If it does violate the rule, the example should SOUND wrong. The part I get confused on is the idea of sounding right or wrong. When a sentence violates OG rules, it ALSO sounds wrong…
DeleteDoes the difference lie in the fact that for UG, the wrong examples are produced to determine what the (unlearned) rules are while for OG the wrong examples are produced knowing explicitly what the rules are? How can they be sure their example sounds wrong because it is violating UG and not just OG?
I can try and answer Anaïs' question to some extent. I'm not sure if any linguist could ever be 100% sure that a contrived ungrammatical sentence sounds wrong because of UG and not just because of OG. But linguists typically model sentences according to particular hierarchical structures, operations on/within those structures, and principles concerning/constraining those structures, and can then construct ungrammatical utterances which simply can't be modeled while still being in accordance with those constraints on possible structures/operations. These utterances would be hypothesized to violate UG. We could always come across a language in which some sequence actually is possible and then would have to revisit our models, and so nothing is 100% certain as I mentioned, but that's the case in any scientific endeavour.
DeleteLike my peers mentioned, I also struggled fully comprehending this detailed reading. However, I will mention what I took away from it and other thoughts about it:
ReplyDeleteFirstly, it is fascinating to see topics confidently explained in other classes be critiqued in ways that reveal a list of misconceptions toward UG. Secondly, this reading builds on previous readings in that it showcases why researchers who are at the top of their fields continue to misunderstand the poverty of stimulus and UG. Thirdly, it may come down to paying more attention to the differences between UG and OG.Additionally, there is a lot of talk about negative and positive feedback. Negative feedback refers to “information that some word sequence is not grammatical.” Positive feedback refers to “information that some word sequence is grammatical.” Moreover, children will only make OG errors.
Based on this, I want to relate to the classes I have taken that focus on our brain forming synapses in learning. Specifically, synapses will be strengthened when we learn something of some value and in constant reiteration, and the stuff we learn that does not hold value or reiteration will fade away. In this sense, would this idea of learning be connected to positive and negative feedback, i.e., learning between a grammatical vs. non-grammatical word sequence?
Maira, yes, P & S point out some errors and misunderstanding about UG – and then they go on to make errors and misunderstandings of their own: What are they? (Please read the other comments and replies for 8a,b and 9a,b).
DeleteDistinguish negative examples or evidence from negative feedback, which is error-correction.
No doubt learning is related somehow to neuronal connections and activations. But that’s been known for decades; what’s new is computational models of how neural networks learn (unsup and sup).
This article as many of my classmates have pointed out is refuting the argument of PoS (Poverty of Stimulus). There is some misconstruction of OG and UG and the differences between them. It's arguing that children lack of positive information of language when that was not the issue that we were aiming to resolve or trying to understand.
ReplyDeleteHelena, what are the issues we are trying to resolve and understand?
DeleteA question about the grammaticality judgments: if all UG non-compliant utterances are also OG non-compliant, how does anyone determine whether they've made a grammaticality judgement on the basis of a learned OG-rule or a UG rule? I'm assuming there is a formal difference between OG rules and UG rules, and linguists producing utterances to test candidate UG rules refer to these rather than their own linguistic intuitions...
ReplyDeleteIn this article the authors argue that the present evidence for the poverty of the stimulus is not valid, because from further examining the corpus, the stimulus is not that in poverty – there is actually sufficient evidence for them to generate certain patterns that were thought to be not generatable from the evidence but can only be acquired by innate language faculties. With the new information technology and building of corpora, researchers nowadays can overturn the judgments that were made decades ago, which proves the usefulness of corpora as people are not going to be making claims basing on their extrapolations on their own experience. However, there is a serious problem on this paper: How much stimulus is ‘sufficient’, or not impoverished? This question is asked at the end of the paper, and the authors propose this as a topic for further research. This must be the most important problem on determining the validity of the poverty of the stimulus argument in our times with readily available corpora.
ReplyDeleteHan, the problem is not positive evidence (UG+), nor how much there is of it; it concerns negative evidence (UG-) and corrections.. The child hears and gets none for UG. And what it hears and gets for OG is irrelevant. Please explain if you understand this.
DeleteHi Professor Harnad,
DeleteI just noticed your response and I am here to explain that I do really understand the concept of the poverty of the stimulus. The poverty of the stimulus is that the children never hear sentences that do not comply with Universal Grammar, like ‘hire universe I wings.’ Instead, they only hear sentences that do comply with UG, like 'I like chicken wings.' So, with a lack of negative evidence, they theoretically cannot learn the language if they learn only by unsupervised and supervised learning. However, they actually do, so the problem arises.
It seems that the paper itself misconceived the concept and I was misled by the paper.
Han, your example was a violation of OG. You can't invent a violation of UG unless you have explicitly studied UG.
DeleteThis reading attempts to question the evidence provided for the existence of a universal grammar by refuting the existing evidence provided for an innate comprehension of specific grammatical constructions. This evidence mainly consists of the argument that children are not exposed to certain grammatical structures enough or at all, but can identify them as grammatical when they appear. So, they must be considered innate to the child. However, the main issue with UG is that children do not make mistakes in UG but a speaker of any language automatically knows when a mistake in UG has been made. Despite this, there is value in this article, as the authors do propose ways in which linguists can be more careful and accurate in their work and outline the interdisciplinary nature of the question. Furthermore, this article makes me question what is fundamentally different about the ungrammaticality of non-UG-compliant sentences, as opposed to non-OG-compliant sentences.
ReplyDeleteElena, both UG errors and OG errors sound "wrong" but we know (or can easily learn) the rules that OG errors violate (and they differ from language to language), whereas we have to study linguistics to find out the rules of UG, which are the same in all languages. More important, we never make errors in UG whereas there are OG errors aplenty.
DeleteIn today’s lecture, we saw how there is no UG negative evidence and children don’t make UG errors. There are no UG errors to be corrected, only OG errors. So, every sentence that an adult, or a child learning to speak, will produce will be UG compliant. The only case when UG- examples are produced is the ones concocted in linguistic laboratories. We can tell when sentences “sound wrong” because it feels like something to us. I wonder if children can make that same distinction right away when they first learn to speak, or if it is a capacity that evolves with the development of their brain. I know that we can have a learning ability encoded in our brain so that we are already able to learn things more easily without it being an innate capacity. But, once children know how to speak, if we were to present them negative UG examples on purpose, could they infer that it also “sounds wrong” to them, or is this ability something that has developed with the evolution of their brain? This could be more generalized to what it means for something to be innate. Is it necessarily something ingrained in our brain from our birth, or could it be a capacity that develops later on, yet is still not learned. It is present in our brain from the start, it is not learnable (it is innate) yet it only activates later on in one’s life.
ReplyDeleteI don’t know if this thought is well explained, please let me know.
Étienne, you are asking how soon children who are learning language can make grammaticality judgments (perceiving OG- and UG- examples as wrong). I don’t know, but I know that as adults (and perhaps already as adolescents) they can.
DeleteI also don’t know whether they can perceive OG errors before they can perceive UG errors, or vice versa, or both at the same time. But I can say that if there is more learning going on between the time children have learned to understand and speak and the time they can make grammaticality judgments, then that ongoing learning can only be OG learning, not UG learning, because they do hear and produce and get correction for OG errors (OG-), but they still do not hear or produce UG errors (no UG-) (hence no correction either). (The POS still rules.)
Another thought I had was regarding the capacity to code innate properties into machines. Indeed, machine learning models nowadays work by taking examples with an enormous amount of data, performing supervised (trial and error with corrective feedback) or unsupervised (passive exposure) learning, and producing output. However, we do not have negative examples of UG, we only have positive examples, and as seen in class with the example of “being awake” positive examples are not sufficient for us to determine the features of something and to categorize it. So what is the software behind coding innate capabilities. Surely there are already ways to do this because we already have Large Language Models such as laMDA or GPT-3 that possess tremendous conversational capabilities and, to my knowledge, they comply with UG constraints. We cannot use classic machine learning techniques because we don’t have the data to give negative feedback to the machine on a trial-and-error task. So I am wondering how this can be coded inside a computer so that, like a human, it appears to have innate UG capabilities and does not make UG mistakes
ReplyDeleteÉtienne, the simplest way a computer program can categorize “innately” is if the feature-detectors for the features that distinguish members of the category C+ from the non-members (C-) are built in in advance rather than having to be learned by (unsup or sup).
DeleteGPT-3 and LaMDA are derived from sampling an enormous number of texts and calculating the frequency of 2-word sequences. (It’s more than that, but this will give the idea, to answer your question.) It generates new text based on such stored frequencies plus some new input text. It doesn’t make UG errors because texts don’t make UG errors (except if written by linguists!) and LaMDA's and GPT-3’s production of new texts is just based on the probability of sequences and co-occurrences in their database. GPT-3 is not a T2 algorithm.
I wanted to expand upon Etienne’s statement, “We cannot use classic machine learning techniques because we don’t have the data to give negative feedback to the machine on a trial-and-error task.” I think that it might actually be possible to use machine learning techniques to discover UG rules because we actually do have examples of non UG compliant sentences (the UG- sentences explicitly designed by linguists). In a previous reply, Prof. Harnad commented that linguists currently go through a process of putting together potential examples of non UG compliant sentences and testing those examples. Provided that there are enough non UG compliant sentences that linguists have already discovered, ML models could be fed these negative samples alongside the massive amounts of positive samples to detect the features differentiating UG compliant and non-compliant sentences. This labeled data would train the ML model using supervised learning. The model could then propose UG non compliant sentences where linguists and non linguists alike could quickly manually verify whether the model is producing UG-sentences. Although it requires an enormous quantity of human resources to verify the model’s results, it takes very little effort for any human to differentiate because it is innate to us. This feedback loop theoretically would help the model get better at proposing UG- sentences and help linguists discover the rules behind UG.
DeleteUpon reflecting on this week’s lecture in relation to this reading, here is what I have understood. Ordinary grammar requires both positive and negative evidence, thus there is no poverty of the stimulus. Universal grammar has no negative evidence whatsoever, thus we have the poverty of the stimulus. In order to learn what things are and are not, why do we need both positive and negative evidence? Having positive instances of something allows us to make the distinction between what is and what is not. If what we need to distinguish lions and non-lions for example are their specific features, you have to find which features are different. You need positive and negative examples to achieve this distinction. And what if we did not have this? In class, we looked at the example of not giving negative evidence. If we are ONLY given positive evidence, saying that humans are Layleks, liberty is a Laylek, animals are Layleks, etc. But if we never define what a Laylek is NOT, how can you get to what a Laylek is? Therefore, you cannot get to Layleks, if there is a lack of negative evidence for non-Layleks. In a more general sense, you need to have the compliment of a category (what does not belong in the category) in order to distinguish what IS in the category.
ReplyDeleteSara very good synthesis! Just a few tiny (almost pedantic) fixes:
Delete1. To learn OG you need both OG+ examples and OG- examples, in order to be able to detect and abstract the features/rules that distinguish OG+ from OG-.
2. Language-learning children never produce or hear UG-, only UG+. So they cannot learn UG (and they don’t need to: they already have UG in their brains when they are born; that’s why they can already produce all and only UG+ utterances almost as soon as they start producing language [triggered by the “trigger” experience, setting UG parameters as they learn OG]).
3. If we never have examples [not “definitions”] of what a “Laylek” is NOT, we cannot learn what a Laylek IS. [Definitions come from grounded verbal instruction; supervised learning just needs L+ and L- examples, to be able to find the distinguishing features.]
4. For any category, C- is the complement [not “compliment”] of the C+ category – that is, all the other categories with which C+ could be confused. With the UG+/UG- category it’s just a dichotomy into which all possible utterances can be split. With the category “apples” the complement (apple-) of apples+ includes bananas+, pears+, grapes+…
Hadrien, both OG errors and UG errors sound wrong to everyone (except very early L1 learners and some early or weak L2 learners). Every schoolteacher knows why the OG errors are wrong, but only MIT (and kindred) linguists know why the UG errors are wrong.
ReplyDeleteAlthough I agree with the poverty of stimulus argument in principle, I think that the argument is only effective when one delves into the supporting evidence behind it; the notion that we must have some device in our heads that biologically makes us capable of learning language so efficiently and quickly does not have it’s best evidence rooted in the poverty of the stimulus. Instead, it makes more sense to look at cases of infants, since it has been shown that despite showing preference to their native language, infants are essentially capably of learning any language before a certain age. If a child hears and speaks English for the first 5 years of their life, then all of a sudden their environment changes such that they only hear and speak in French, they still will be able to learn French with native fluency, suggesting that language acquisition must be innate in some way.
ReplyDeleteBrandon, see the comments and replies about critical periods and first-language (“native” language) and later language proficiency (L2, L3…) as well as UG parameter-settings.
DeleteYes, some people can get near-L1 proficiency with L2 and more. The controversy (or uncertainty) is about whether they get equal proficiency for detecting UG- in their L2 or later. The question is about the parameter-settings on UG. There is only one UG, and each language sets some learned (OG) parameters on UG. If there’s just one L1, no problem. But if L1 and L2 have different OG parameters for L1 and L2, what are the parameter-settings on UG? Are there two different settings? Do they interfere with one another? Are grammaticality judgments about OG violations (OG-) in L2 as accurate as judgements about OG violations in L1? (And what about when two different L1s are learned at the same time?)
The principal supporting evidence for UG and its unlearnability and innateness is the child’s capacity to produce all and only UG+ utterances in L1 without ever hearing or producing UG- (in L1).
For the stout-hearted I've added two more supplementary references on UG after the P & S skyreading at the top of this page.
ReplyDeleteRosalie, but the main reason why UG cannot be learned without negative examples is that it is complex and highly underdetermined. OG is simple, and much less underdetermined. (What does "underdetermined mean"?
ReplyDelete"What does "underdetermined mean"?"
ReplyDeleteThe underdetermination of a theory refers to the idea that the evidence available to us at a given moment is insufficient to determine our belief in response to said theory. For instance, if a positive correlation is found between the number of sales of firearms and murder rates, it could be that the rising number of firearms circulating in a community is causing the rise in the murder rate. But it also could be the case that the rising murder rate (that is not necessarily due to gun violence) is encouraging people to purchase firearms for self-defence. In this case, the positive correlation between the number of firearms sold and the murder rate is evidence that underdetermines what we should believe about the causal relationship between the two variables (the positive correlation may be due to a third factor, i.e., socioeconomic inequality).
In UG's case, UG is underdetermined because children arrive at grammatical theories without negative evidence (which is required for category learning). OG is learned and is learnable in the same way as any other category learning. In other words, the abundance of positive and negative OG examples in a child's environment allows them to pick up OG through either observation (unsup learning), trial-and-error/correction (sup learning), or pure verbal instruction. The sufficiency of positive and negative evidence of OG in one's learning environment determines the reliability of the relationship between the grammatical theories a learner arrives at and their ability to speak/write grammatically (OG = determined = a form of category learning facilitated by exposure to both member items and non-member items).
On the other hand, children and adults do not make UG mistakes. In other words, negative UG examples are absent in a learner's environment (POS). That said, children are UG compliant from the get-go, despite not being exposed to any UG mistakes. The insufficiency of negative UG evidence underdetermines UG theories in the sense that we are not sure how exactly they came about. Since negative UG examples are absent in our learning environment, it is impossible to distinguish positive UG examples from negative ones. Namely, a positive correlation between OG-compliance and the ability to tell if a sentence "sounds right" determines what we should expect about the causal relationship between the two because the former represents an attainable category learning process involving exposure to positive and negative evidence. However, the positive correlation between UG- compliance and the ability to tell if a sentence "sounds right" does NOT determine what we should expect about the causal relationship between the two because the former is NOT a learnable capacity. Strictly speaking, we cannot be confident whether it is the capacity to notice a phrase "sounds right" precedes UG or vice versa, in a similar logic of how the causal relationship between murder rates and the number of firearm arms sold goes underdetermined if there is no available third party factor.
Yucen, chess has rules. If you know and use all the rules of chess, you will only make legal moves in chess. The rules of chess are usually learned from verbal instruction. But they could be learned by supervised trial/error/feedback learning too, with someone telling you “legal” or “illegal” right after you make a correct or incorrect move. The rules could also be learned by unsupervised learning and positive-only evidence, if you watched enough chess games and saw what moves are made (and what moves are not made).
ReplyDeleteSo the rules of chess are not underdetermined. They are fully learnable in all three ways.
The rules of OG are also not underdetermined; they can be learned in all three ways.
The rules of UG can probably all be learned by adults through verbal instruction (given by linguists).
The rules of UG can be learned by adult linguists through supervised trial/error/feedback learning, but only because they already have UG in their heads, giving them feedback about what sounds correct and incorrect as they try out different rules by trial and error. So the rules of UG are not underdetermined for adult linguists, but that is only because they have the supervisor already built into their brains innately.
The rules of UG cannot be learned in this way by the child. So the rules of UG are underdetermined for the child. The child must first learn language, by hearing it, trying to speak, and getting some feedback on OG rules (which can be learned, like chess, by unsupervised or supervised learning). Meanwhile, the child makes no UG errors, so UG must already be in its head from birth, innately, and “triggered” or activated by starting to learn language, including OG.
UG rules are underdetermined by the evidence available to the child, hence UG must be innate, because the child can nevertheless comply with them. OG is not underdetermined. Learning can determine, by trial/error/feedback testing, what are the rules of UG.
The reason the causes underlying the correlation example you gave (gun sales and murder rates) are underdetermined is because you can’t (i.e., you are not allowed to) test the possible causes separately by experiment. But with another problem of the same form as the underdetermined gun/murder problem – for example, whether plants grow better in the summer because of increased sun or increased temperature -- the causal explanation could be determined by trial and error in a laboratory. So the underdetermination can be resolved.
The underdetermination of UG was presumably resolved by prior evolution (which is a form of supervised learning – supervised by the adaptive consequences of genetic variants), but it is not clear how.
I appreciate the critiques and different point of views that this reading takes a stance on, I did however find that it took a while for me to actually feel as though an argument was being made. Pullum, for the first significant part of the reading, rightfully critiques and questions mainstream cognitive and linguistic beliefs, but I found that it took a long time to come across a new proposition or an actual idea of what their beliefs on the topic might be.
ReplyDeleteThis paper and this discussion definitely hammer home the point that UG must be innate in some way, as children do not make UG errors. It was mentioned briefly above that children would not repeat UG errors made because they would never hear them, unless their parents were linguists and making them on purpose. That does make me curious on what the effect would be on a child that is regularly exposed to say, their parents intentionally making UG errors at home, while simultaneously going through the same positive and negative evidence of OG at school and other environments. Would their understanding of OG or UG be warped in any way, or would the innateness of UG and the regular process of language learning they’re experiencing outside the home override the UG errors they’re exposed to?
ReplyDeleteThe researchers here do not adequately distinguish between UG in OG when they set out to disprove the POS argument. We know that UG must be an innate feature in the human brain because when children are learning language, they only produce UG+ utterances while never having heard UG- utterances. Since you need to be exposed to UG- utterances in order to learn what distinguishes them from UG+ utterances, and children are not exposed to UG- utterances yet still show a complete understanding of UG rules, the only conclusion to make is that UG is an inborn feature. The article does not really address any of these points.
ReplyDeleteThe argument in this paper does not establish what UG or OG is and ignores how the environment heavily impacts OG. Various examples used to argue against POS used aspects of OG, like dialect differences, not UG. The argument for POS is that children must be exposed to a large amount of data within their environment to acquire language successfully.
ReplyDeleteThis paper reminded me of a sociology class I had taken on neighbourhoods and inequality. In particular, an article we had in that class focused on how language changes due to underfunded schools, but slang and other developments in language, like AAVE, still follow fundamental grammar.