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.
This article sets out to show that proponents of the "Argument from Poverty of the Stimulus" argument (APS) have yet to produce evidence that backs up their theories. Of the evidence that they have produced, all can be refuted using corpus analysis techniques that were previously unavailable or ignored by researchers. Notably, Pullum does not aim to disprove this version of nativisim with empiricism: rather, he uses empirical evidence to demonstrate that the nativist position isn't as well founded as it purports to be.
ReplyDeleteThis is the scientific method at its best: while evidence can be used to disprove or bolster theories, it can also point to a gap in our knowledge. Pullum leaves us unsatisfied with the APS, and in doing so, he pushes us and future researchers to investigate the theory further. Crucially, he encourages his peers in generative linguistics to adopt strategies to which they were traditionally antipathetic, like mathematical learning theory and corpus linguistics.
As Pullum demonstrates, research's value isn't just in proofs and theories; it's in identifying the gaps in our ideas so that we can bridge reasoning and reality.
Don't get too enthusiastic! Pullum does the same thing Pinker did: he ignores the distinction between UG and OG, thereby begging the question.
ReplyDeleteAll the violations Pullum finds in his big-data text corpus are just violations of OG, not UG. It remains true that no one -- child or adult -- hears or produces UG violations, let alone corrections... except adult Chomskian linguists. But they are doing it to try to reverse-engineer the rules of UG, by trying out potential rules to see if they work. And their corrective feedback for the UG-violating utterances comes from their own innate UGs!
It turns out Geoffrey Pullum (who is quite clever) was not quite right about the eskimo snow term hoax either.
Pullum, Geoffrey K. The great Eskimo vocabulary hoax and other irreverent essays on the study of language. University of Chicago Press, 1991.
Lupyan, G., & Zettersten, M. (2020). Does vocabulary help structure the mind?.
This paper seems to apply poverty of the stimulus only to positive feedback information gathered from a child’s environment. The authors supply strong support to suggest complex and specific grammar principles are relatively commonly heard by children during language acquisition. So, it is probably not correct that since children aren’t exposed to these examples, they must be innate. However, the authors’ claim that lack of positive evidence for acquired grammar is the strongest support for nativism seems to be a matter of opinion. The authors set aside the arguments of incompleteness and positivity by failing to examine the lack of UG-violating examples in speech. As was discussed in categorization, to learn rules and categories, it is crucial to not only know what is inside a category but also what is not. Pullum and Scholz show that what is inside a category (in this case correct grammar) is supplied to children, but don’t address how children know what is outside the category. This is not to say that children don’t make errors, but they don’t make ones that violate UG. How would they know to only make OG errors and not UG if not for some innate knowledge of UG grammar?
ReplyDeleteWhat's not innate is what can be and has been learned by the child through trial and error, with feedback from correct and incorrect responses (supervised/reinforcement learning). Trial and error requires both positive and negative data (i.e. both members and nonmembers of the category, otherwise it is the "Laylek" situation). Pullum's survey of the data available to the child does not distinguish OG -- for which there is plenty of data, both positive and negative -- and UG, for which there is only positive data (heard or produced by the child).
DeleteIn terms of useful takeaways from this paper I think the biggest one is that children do receive a richer and fuller set of inputs than I had previously thought. The authors are self-aware that they have collapsed the acquisition of language by children as presented on pages 12 to 13 into just focusing on all of the positive stimuli inputs children would be expected to hear. Fine – but after reading some of the other skywritings and replies for this week I really do not think that a lack of negative examples can be made up for by this abundance of positive examples. It is great to show that children are exposed to unique and what were presumed to be very rare formulations, but I am really not sure that this pokes as many holes in the nativism theory as they suggest.
ReplyDeleteIn connection to the idea of a purely correlational learner that was brought up in the Pinker paper, I am not sure I think having this large data set as discussed in the Pullum and Scholz paper is a solution to learning language. With no innate filter to parse the input data into these different phrase structures and have a hint as to what features should be remembered because they are important, how would all of these utterances be stored in the child’s brain so that different correlations could be extracted? Our memory is good for learning thousands of vocabulary words, but this seems different to me than storing entire utterances in order to learn grammar. I hope that this is not the equivalent of a granny argument but it is something that I am curious about.
A lot of OG can be picked up passively, by unsupervised learning. But this does not work for UG, because of POS, no matter how much the child hears. To do it by supervised learning, they would have to make UG errors and be corrected. They don't, so they aren't. They could be read every sentence in Pullum's book, and go on to hear the entire database, day after day; that still won't provide the negative evidence they need in order to be able to learn to produce and recognize all and only UG-compliant utterances -- unless they can do it already.
DeleteAs a complement to Prof Harnad's reply, I'd like to add that it is (presumably) UG that helps us make sense of the data we hear, therefore bypassing the POS. Pullum & Scholz are not suggesting that having a complete data set of everything a speaker hears throughout their language learning (i.e. what they call a "full input corpus") would help a language-learner acquire, well, language. They are simply stating that having this input corpus might help us (i.e. scientists, researchers) determine, finally, what children do and don't hear in their input.
Delete*what children do and don't hear during language acquisition.
Delete(Of course, this would require a full input corpus from multiple different languages and multiple different OG environments within each of those languages to provide anywhere near sufficient evidence for generalizing what children do or don't hear).
UG cannot be learned through unsupervised learning (i.e., passive exposure), no matter how much you are passively exposed to. And this is equally true of the child and the adult (including the adult linguist!).
DeleteLearning UG (if you don't know it already) would require supervised learning: active trial-and-error with corrective feedback; trying to do the right thing.
The way adult linguists have learned UG has been by pondering UG violations (the starred sentences like *"John is easy to please Mary", which they have conjured up (I'm not sure how) and trying to guess why they are wrong. (The fact that they are wrong is signalled to linguists by their own brains, their own UG.) But even with feedback that the starred sentences are wrong, linguists can't abstract UG; they can only say that the reason they are wrong is not explained by OG.
What linguists have to do to abstract the rules of UG is to hypothesize them explicitly, and then apply and test their hypothesized rules, to see whether they do succeed in generating UG-compliant sentences, and no starred sentences.
In other words, they have to do category learning, and with explicit hypotheses about what are the features (rules) that distinguish the members from the nonmembers of the category "UG-compliant." (This is done by teams of linguists, across time, so the task gets easier as the UG-structure they succeed in extracting (like sculptors) grows.
I’m going to use these quotes from 9a because I feel like they tie in pretty nicely
ReplyDelete“The girl giggled and Don't giggle me… unless their parents slip them some signal in every case that lets them know they are not speaking properly, it is puzzling that they eventually stop”
“If the world isn't telling children to stop, something in their brains is, and we have to find out who or what is causing the change”
This sounds familiar. I’ve definitely heard people with English as their second language say “I’m going to learn you” instead of I’m going to teach you. The authors say that something seemingly innate eventually stops children from making these mistakes, so are they implying that these are UG mistakes? And if yes, why do my friends who (I think) learned English as a second language continue to make these mistakes? Are we more sensitive to UG during the critical period? My understanding was that nobody, children nor adults, would be making UG mistakes (except linguists that are purposefully looking to make those mistakes) so this sort of confused me.
But back to this reading, I think we all agree that learning OG requires some form of positive and negative feedback. After all, to know what something is, we must also know what it is not and kids do this by making mistakes and being corrected. However, with UG, kids end up learning the right thing although they have never been exposed to the wrong thing because UG mistakes aren’t made. To make progress towards the reverse engineering problem, this distinction definitely needs to be made because UG needs to somehow be hard coded into our T3 robot while OG would probably be learned (I’d imagine through supervised learning). Trying to teach a robot UG through supervised learning, if it works, might give us weak I/O equivalence because we pretty much know for a fact that we don’t learn UG. The only benefit I’d see in this is that if we can’t figure out how to hard code UG then we can teach them to the robot instead. We know that that robot won’t be learning language the way we learn language but if we manage to teach it language at all that’s a win. But if we are going to spend time coming up with enough wrong examples of UG to teach the robot, we could just hard code them in instead and save training time.
I wouldn't say that "I'm going to learn you" is a UG mistake. It's an OG mistake in English, but I don't think that there are any UG parameters that make "to learn" ungrammatical here. I'm unsure which dialect of English this is from, but I've definitely come across "I'mma learn you something", which might not be grammatical or acceptable in all English dialects, but does occur without, presumably, violating UG. I also don't think we are "more sensitive to UG during the critical period". We don't know UG (explicitly at least) at any point in our life, but we still respect it.
DeleteL2-speakers (or even L3-, and so forth) of English make OG mistakes like "I'm going to learn you" because, presumably, it's a perfectly 'legal' way (according to their L1's OG) to communicate the intent to teach you something. That also means it does not violate UG, because otherwise that would be 'illegal' in all languages and never spoken by L2+ speakers.
Lyla, Pinker was simply failing to distinguish UG from OG in these examples, and that is misleading, and leads to misunderstanding. (That is why he was not a reliable oracle to the psycholinguistic community on what UG and POS were really about.)
DeleteYes, there is a critical period for 1st-language-learning (L1) and for UG parameter-setting. After the critical period, some kinds of errors will persist in 2nd-language (L2) learners. But not many, because even with wrong parameters, all languages share most of UG.
(I doubt that L2 errors handicap their users enough to be called “maladaptive”: do you think they do? Do they really reduce survival, reproduction or even career or social success? In our tribal, territorialist past, though, they might have been cues, along with appearance and other features, as to who were from your own tribe and who were potential aliens. That’s probably also why we learned to get so good at overcoming and concealing our speech errors.)
Yes, “something innate” prevents the child from making UG errors — and what stops them is UG iyself, which they already have, rather than learning it, as they learn OG.
If (as it seems, according to the evidence) UG is innate, then it is not relevant (to reverse-engineering human cognition) to teach a robot UG, unless we are trying to model the behavior of adult linguists, for some reason!
For child L1 learning, according to UG, the child would not be able to learn language at all if it did not already have UG. So it’s not clear why or how to do what you are imagining. It’s not about storing all the errors! It’s about how to design the robot to generate any of them at all.
No, if UG is right, it is the linguists who have already done the reverse-engineering. The UG algorithm would need to be built into the T3 robot from the beginning, except for parameter-setting.
Eli, you are right about “learn you”: it’s just an OG error. (This habit of treating OG errors as if they were UG errors — and as if that was what UG was about, instead of studying UG — has been responsible for a lot of the misunderstanding as well as the utter nonsense that has been sputtered by “critics” of UG who have no idea of what they are talking about!)
As to whether we (whether children of adults) “know” UG: we know it implicitly: our brain knows and is constrained by it; we just don’t know it explicitly — unless we study Chomskian grammar!
About learn/teach: they're obviously closely related semantically, hence often morphologically too (all OG):
G: lernen/lehren
F: apprendre/apprendre à
R: uchit'/uchit's'a
H: tanulni/tanitani
A?: taealam/yaelam
@ Eli "learn you" originates from AAVE (African-American Vernacular English) I believe
Delete@Professor I'd like to explore a bit on what you stated about T3 and UG: "the UG algorithm would need to be built into the T3 robot from the beginning, except for parameter-setting"
DeleteCorrect me if I am wrong, but this proposition that you stated seems to imply that the T3 robot would have UG capacities/algorithms built in it, without a pre-set language. Even if it is the case that we would set the Robot's "innate" language to English, how long would the robot have to be functioning in the real world to determine "indistinguishability from a human being" in terms of its linguistic capacities? Would it have to follow the error-ridden pathway of children figuring out how to use their 1st language, or simply properly acquiring a language with the absence of negative evidence?
My hunch on this question is that it would be the latter, as that is what has been discussed the most and also, we wouldn't have a T3 robot with a developing "brain" but one that has "adult" brain capacities (including linguistic ones).
To follow up on that, according to UG principles, this would mean that a T3 robot that has had UG successfully implanted in it would be able to learn any language with the same degree of success/difficulty as any other language. Does this mean that to determine whether or not we have successfully figured out UG, it's not sufficient enough for a robot to just have learned one language, but maybe build a replica and have it learn a different language?
Remember that T3 is about capacity. T3 has to have the generic cognitive capacity of a human. This includes the capacity to learn any language as a child -- or if the T3 is an adult and already has a 1st language, then it has to have the generic human capacity to learn a 2nd language. (That capacity varies a lot across people, and the differences don't matter for TT.)
DeleteYou can answer all your own questions if you just remember that TT is a test of capacity. The child vs. adult issue is moot because we are too far from anything remotely approaching TT scale to worry about whether not only learning capacity but developmental history is necessary for passing T3. And nothing we've learned about UG is relevant other than that T3 needs to have generic language and language-learning capacity.
Pullum takes the following to be the correct version of the POS argument:
ReplyDelete"There could be evidence for a certain grammatical rule in a language that could be found by an adult linguist through intensive search but would never appear frequently enough in conversational data for a child to find it."
After reading this article, I came to the conclusion that the four arguments in section 4 that supposedly undermine the APS pertain not to UG but to OG (which is confirmed in a comment above by Prof. Harnad). There being obscure grammatical rules in a language that children know or do not know because of lack of exposure is not what the POS argument is about. The POS argument is really about language rules that everyone knows, universals, whatever language they speak. It is that capacity, unique to humans, to learn any language as a child. The POS argument pertains to the fact that we learn and master any language as children much faster than would be expected if we did not have any prior knowledge given to us by innate language learning mechanisms.
On the other hand, I think that the empirical methodology outlined by Pullum might be useful to define UG more precisely. I don't know a lot about linguistics, but from what we have read so far, I am getting the impression that no one really has clearly defined the rules of UG and its structure. Most of the time, it is mentioned in quite abstract terms. Although defining it as "parameter setting" and potential "constraints on thought" is interesting, I still don't know what the parameters are (or if anyone knows them) or what those constraints might be. Potentially, those constraints prevent us from figuring that out too.
"I don't know a lot about linguistics, but from what we have read so far, I am getting the impression that no one really has clearly defined the rules of UG and its structure."
DeleteThis is incorrect, and can only be remedied if you actually study UG.
(Chomsky often pointed out that such statements would never be made about differential topology by someone who had never studied it. But since we all speak language, we imagine we have enough insight to form a judgment about UG without studying it...)
Hi Solim, I like your comment!
DeleteI am also having trouble thinking about rules of UG and understanding what they do and what they mean exactly. I am sure that this is due to the fact that I am not studying Chomskian linguistics as professor Harnad answered.
What we know:
UG rules do exist; from what I understand, they are just rules that linguists derived from studying UG violations (violations that they came up with/ invented). UG rules are rules for generating UG compliant strings. They would not be learned by children due to POS. So they are innate in some way.
Now, I think I understand what bothers you. Correct me if I'm wrong, but I may also share this concern with you. What does it mean that the rules are innate? What sort of implementation would that be? Maybe innate rules of UG are just how we think (and thus not "rules", but information-processing or categorization). Rules of grammar are for linguists... They look like abstractions and they are explained in words. In my opinion, those sorts of rules are far from giving insight into cognitive mechanisms underlying them.
I would like to hear what professor Harnad as to say about this.
There is no reason why grammatical rules could not be innate, but Chomsky has suggested that they may be side-effects of innate constraints on thinking. The problem in the case of UG is to explain how and why UG evolved: what was its adaptive advantage?, and how could it evolve gradually?
DeleteIt seems clear to me that UG's adaptive advantage may be the fact that it restricts the "legal" linguistic variations present in language. The benefit is, of course, that every variation of OG still respects some, theoretically, fundamental rules and is therefore not unlearnable to wielders of very different OGs. As to how it could have evolved, it also seems quite clear to me that those who had similar understanding of language could communicate and therefore survive and even thrive in groups. Those who couldn't communicate were potentially eventually separated from the in-group. If we subscribe to "thought drives language", then UG could have evolved gradually when individuals who could communicate well with their group survived, passed on their genes to their children, who were born with a smidgen more predisposition towards a "standard" language, who then used it "better" and passed that down to their children, etc. etc., until it became the innate UG we theorize today. These all are, of course, full to the brim with hypotheticals. But the maybe remains.
Delete“What does it mean that the rules are innate? What sort of implementation would that be?”
DeleteMaximilien, I think that I also share your and Solim’s questions regarding what the implementation of UG concretely looks like. Professor Harnad answered that there is no reason why grammatical rules could not be innate, but I still have trouble conceptualizing how exactly this would work.
I suppose that the element that seems counterintuitive is that there can be constraints on something that must be learned (language) before it is even learned. When I think of other skills that must be learned, such as playing guitar, for example, it seems unimaginable that that there would be built-in rules for how I execute this skill.
I guess that something like walking would be a better parallel, though. Although we must learn to walk, as we learn how to speak a language, it is such a fundamental function that it is inevitable that all children will learn to do it (unlike playing guitar), and there is therefore already an innate structure for what it is we are learning. If we were inventing walking or grammar from scratch, we would see a lot more diversity across cultures in how this is executed.
Either way, UG is mind-blowing!
In "Empirical assessment of stimulus poverty arguments," Pullum understands the Poverty of Stimulus argument (POS) as referring to the fact that children are not exposed to data about the structure of their language which would be necessary to learn that structure; therefore, the structure must already be there -- that is, it must be innate. Pullum argues that empirical evidence sheds doubt on the POS. One example he gives concerns variations in plurals of compound words across linguistic cultures. He notes that in the USA, they would likely say "a drug problem," whereas in the UK, they would say "a drugs problem." Pullum argues that this variation would not be possible if there was an underlying UG which would put a constraint on which compound words could or could not be plurals.
ReplyDeleteAs has been noted in this thread, the problem with this example -- and Pullum's argument more generally -- is that he addresses variations or incorrect formulations in ordinary grammar (OG) but not UG. UG concerns the internal structure of sentences. In the case of "a drug problem" vs "a drugs problem," the structure of the formulation is the same (it's just that in one case a plural is added to a word and in the other it is not), indicating that this variation is not in fact a variation in UG but in OG. For Pullum to make a better argument, he would have to show that variations exist in the internal structure of sentences which go against the principles of UG -- indeed, he seems to think that the POS concerns a supposed lack of data about the structure of languages -- so it is strange that he uses an example which does not in fact have to do with the structure of the formulation.
Any single violation of UG (or UG) can always be adopted as an idiom. That just leaves an infinite number of other violations that would also each need to be adopted as an idiom, one by one...
DeleteI agree with others when they say that the definition of UG has been fuzzy. I think it is interesting that Pullum’s empirical approach relies heavily on determining If examples can be explained by data-driven learning which I think is how he ends up discussing OG rather than UG. In fact, can someone be able to explain in kid-sib fashion why the authors believe, “until data-driven learning is investigated in more detail, linguists will remain ill-equipped to do more than fantasize and speculate on the matter” (p. 47). Is it because if you can prove something can be learned through data-driven learning then it cannot be a part of UG? How then can we determine what UG is if we do not learn it? Also, why can we not simple find common syntactical characteristics of all languages and call those part of UG? Surely linguists have a pretty good understanding of what is shared by all languages at this point.
ReplyDeleteOn the fuzziness of UG, see the reply to Solim, above.
DeleteIt has already been shown that UG cannot be learned on the basis of the data heard and produced by the child.
What are the "common syntactical characteristics" that forbid "John is easy to please Mary" (and infinities of other starred pseudo-sentences)?
UG cannot be abstracted by unsupervised learning.
In this article, Pullum questions whether nativists have really done their due diligence to empirically show cases of APS (or POS is you prefer). He proposes a structure to provide empirical support for APS:
ReplyDelete-Acquirendum (essentially what is the rule or principle we are trying to say is innate)
-Lacuna (a set of sentences that can be thought of as the data one would need to acquire the acquirendum via learning)
-Indispensability (argument that one could not acquire the acquirendum without the lacuna sentences)
-Inaccessibility (show that the lacuna sentences are not adequately accessible)
-Acquisition (show that children do acquire the acquirendum during childhood)
The inaccessibility evidence proves to be exceedingly difficult to show since the thresholds for inaccessibility are not strictly defined. How often would a child need to encounter a lacuna sentence for them to learn the grammatical principle? How rare do the instances have to be in order to claim that they're so inaccessible that a child could not possibly have acquired acquirendum with that level of exposure? I suppose it ends up being a moot point, since as many have pointed out, Pullum does not make the important distinction between UG and OG. And for UG, there is not an acceptable level of negative evidence since the entire APS/POS argument is that we have UG in the complete absence of any negative evidence. We've actually seen quite a few papers in which the author does not make a distinction between OG and UG. Is this because we have only recently separated the two into distinct categories? I also wonder what approach we can take to reverse engineer UG and to define its principles. It clearly needs to be separate from OG but this seems to be a very difficult categorization problem to tackle.
See the description of how linguists have been abstracting the rules of UG by hypothesizing and testing rules.
DeleteFor the rest, yes, it's largely due to (1) failing to distinguish UG and OG and (2) fantasizing instead of mastering UG technically.
Seeing the repeated critique of Pullum and Scholz disregarding the distinction between UG and OG, It does make me wonder if this paper would have even been possible in its analytical form had this distinction been acknowledged. With such a strong emphasis on empirical data, would the paper not have been done in(looking past the fact that Pullum's examples end up only including OG errors) by attempting to distinguish between two things that we cannot differentiate between in any empirically satisfying way.
ReplyDeleteHi Deidre,
DeleteI personally think that Pullum could have used empirical data to discuss APS (though I don’t think he would have been able to argue against it but instead, support it if he had done so), but we would have to do it in a very specific way. One thing that bothers me about this paper is that it seems to me very ‘anglocentric’ in that all of his examples are centred around specific aspects about English’s grammar that are obviously instantiated (quite possibly because of discrete parametric changes) quite differently in other languages, e.g. some languages that are geographically close and linguistically related don’t invert their auxiliaries to constitute questions. Of course, the medium of the text is in English, so I’m not saying that English can’t be the main examples that he would use, but in order to prove APS, Pullum should have studied grammar properties that are universal and provided examples in English and in a language or languages that differ from English.
In Pullum and Scholz' article, the authors describe four different cases that they thought provided support for the APS:
ReplyDelete1. Plurals in noun-noun compounding: Children will use an irregular plural as the first element of a noun-noun compound like teeth-eater. (page 24)
2. Auxiliary sequences: Children know the MHBV sequence. E.g., You should have been attending to the lesson. (page 28)
3. Anaphoric one: Children know the correct use of the word "one" when it refers to another word. E.g., This box is bigger than the other one. (page 32)
4. Auxiliary-initial clause: “Children know the auxiliary-initial positioning in polar interrogatives in languages like English and Spanish.”
E.g., The dog in the corner is hungry. → Is the dog in the corner hungry? (Children move the auxiliary "is" to the front.)
E.g., The dog that is in the corner is hungry. → Is the dog that is in the corner hungry? (Children move the "is" in the proposition to the front and drop the other "is".) (page 36-37)
Based on our current understanding, we know that these examples apply to OG, not to UG.
As well, we know with confidence that:
- Humans are genetically more motivated to learn language than other species.
- OG is a set of learned grammatical rules for a specific language, and UG is a set of innate grammatical rules for all languages.
- There is no negative evidence for UG. We are only exposed to UG-compliant examples.
- There is negative evidence for OG. Children and adults who make OG mistakes can learn to correct them through feedback from supervised learning.
To be honest, I have been hesitant to write my skywritings for this week because I still feel a bit lost. I know that word order and recursion are under UG. Are these the only examples of UG we have right now? Can anyone provide some other examples?
I believe UG also dictates which kind of grammatical markers can occur and where in the sentence they can occur, though don't quote me on that.
Delete(Also, we can't claim that humans are more motivated to learn language than other species - we can't know that this is true, or that motivation was the catalyst for us developing language. As discussed in readings and lectures, propositions seem to be at the root of the development of language, but we don't know much else. Food for thought: might propositions and UG be in some way linked? If we subscribe to "thought drives language", could it be that the way that we are built drives our ability to make propositions, which in turn shaped what is and is not legal according to UG?)
@Eli: your food for thought brings to mind the weak sapir-whorf hypothesis, which says that language can influence how we view the world. In another comment below, Matt also brings up a comment about "unthinkable thoughts" and that maybe UG sets some sort of limits on what can be considered "thinkable". As everyone has already mentioned, it might be this false sense of confidence about UG that makes people like Pullum forget that there is a distinction between UG and OG.
DeleteAnd @Ting, I would say that UG isn't necessarily a set of innate rules, but that they are also just a set of possible parameters that can be set to learn any language, because every language exists somewhere within the bounds of those parameters.
Commenting on the youtube video:
ReplyDeleteMy understanding of UG was that there were some rules underlying all languages which we know innately. We do not need positive stimulus to learn it, we do so in the absence of stimulus. However, I am confused about the example you used of Gaddafi arbitrarily making a statement (which we know to be incorrect in English from UG) legal, following which it would become part of UG within a few generations. Does this mean that UG changes with time? If so, if a statement that used to be illegal in one language becomes legal, does that mean that this kind of statement becomes legal in all languages?
The issue is not the absence of positive evidence but the absence of negative evidence.
DeleteOG is always changing, so yesterday's error may become today's correct usage.
UG does not change, but there are an infinity of potential positive (UG-compliant) and negative (UG-violating, starred) sentences. Any individual UG-violating sentences could be adopted as an idiom: We can all agree that *"John is easy to please Mary" just means "It is easy for John to please Mary" -- but that leaves completely unaltered the rest of the infinity of examples of the underlying UG rule (whatever it is: neither of us knows what it is: only those who have studies the technical details of UG know).
I have in my notes for the lecture from this week that some things are unthinkable because it violates UG. If this is the case, how can we have negative examples of UG if we cannot think them? Maybe I am taking this note too literally. If UG is innate or inborn then why would negative examples even exist? You write that an individual UG-violating sentence could be adopted as an idiom; would this mean that any UG violations could be interpreted as having meaning? Could this not then turn negative UG examples into positive UG examples?
DeleteIn response to Matt, UG is supposed to be a fair reflection of what you can think, a constraint on thought. So there are no negative examples because we simply cannot think these sentences/situations. The reason why UG isn't learnable, whereas OG is, is because there isn't this type of feedback with negative examples. As Harnad said earlier, these examples of UG-violating sentences were produced by linguists. No human or child naturally thinks "John is easy to please Mary." For the case of the idioms, they are isolated UG breaking examples and I think you can say that they are UG violations that could be interpreted as having meaning.
DeleteWe know from our study of categorization that you cannot create categories without negative evidence. This is essential to the argument from poverty of the stimulus, or APS, which states that there exists in children some innate linguistic nativism because despite a lack of negative evidence, children obey Universal Grammar (UG). It is important to distinguish UG from Ordinary Grammar (OG) because children make OG violations all the time when they are learning a language. However, these errors are sometimes corrected, and if not, children can eventually pick up OG from unsupervised learning. In contrast, from a very early age, children obey UG without hearing any negative evidence. Where many of the arguments Pullum analyzes in this article fail, and where Pullum himself fails, is this important distinction. The fact that children make OG mistakes is not an argument against UG. Furthermore, Pullum appears to misunderstand the APS - “All we will say here is that for those readers who are disappointed that we will not treat the negative data issue further, and feel that the APS we do discuss is not the one that really deserves the name, we can only suggest that they reinterpret our abbreviation “the APS” to stand for “the Argument selected by Pullum and Scholz.””. The author eschews negative evidence and instead focuses on the amount of positive evidence, which is not what the cornerstone of the argument is. Universal grammar is a mystery because there is an absence of negative evidence, not an absence of positive. UG must be innate because without negative evidence, there would be no way to create the categories.
ReplyDeleteI'm not sure if children could pick up OG from unsupervised learning. As you said, OG requires there to be both positive and negative instances. I'm not sure that by merely hearing the negative instances, the child (or adult who is learning a second language) could tell that it is wrong. But you're right that Pullum (like Pinker in the previous paper) fails to distinguish a difference between OG and UG. While they focus too much on the positive evidence in the environment, it is the lack of negative of evidence in the case of UG that is curious, since the child can still learn "correct" sentences, thus have an innate mechanism for categorization. I wonder if this is facilitated by feeling, and that lazy evolution has allowed us to "feel" something when we produce a "right" sentence using UG.
DeleteI think unsupervised OG learning is a very interesting point because it should technically be possible (since there have been natural language processing models that use unsupervised learning and can parse and respond to speech) but I don't think it could feasibly be the case for us because it would require so much exposure for us to learn OG in a completely unsupervised way, if it's even possible (although I also don't think it's possible for us to attempt to learn OG in an entirely unsupervised way since we are surrounded by people that will, even if they don't outright correct us, react to the OG violation in some way). Even though I think maybe eventually we would be able to determine negative instances simply by hearing them because we've heard so many instances that we've developed a certain understanding of what it should sound like, I don't think this can apply to people due to the short time-frame and limited exposure.
DeleteI think the point about it feeling like something to produce a right sentence using UG is very interesting, especially in relation to Matt's point about the impossibility of thinking a sentence that violates UG.
It's a very interesting point and might explain why we never encounter UG violations, especially if the feeling of wrongness is innate, so we don't say them as children because they feel wrong and soon we stop thinking them at all because of the feeling of wrongness. Although it might be a part of it though, I don't think it could explain all of UG and why we never think UG violations because, for me, at least, OG violations feel wrong and uncomfortable too and yet I still think them with some frequency (although I don't know if this might be a side effect of English being vastly different, grammatically, from my native tongue.)
Hey Wendy — I liked your idea that this phenomenon (the observation that children don’t make errors that violate UG) might perhaps be facilitated by feeling. This topic of feeling has come up time and time again throughout this course. I wonder if the fact that children don’t make errors that violate UG somehow fits into the hard problem of how and why we feel a certain way, when we do things. Perhaps the “why” of the hard problem is that feeling makes survival and evolutionary processes simpler and more efficient. In terms of language, having this seemingly innate “feeling” that certain sentences are wrong, would streamline a child’s process of learning a language, which is a critical element of our species’ evolution and survival.
ReplyDeleteIn addition to using Wendy's idea about feeling to explain why kids don't make UG mistakes, I think feeling can also potentially be used to explain why linguists can come up with UG non-compliant examples, based on their feeling of UG, as well as their explicit knowledge of many linguistic categories and examples of UG compliance, but cannot generalize those examples into rules. Feeling that something is UG non compliant seems in this way a little bit like Searle's feeling of understanding English, and not understanding Chinese. The feeling that something is UG compliant or not is a finite feeling, it does not correspond to or arise from a set of rules which can broadly be taught (otherwise it would be easier to teach).
DeleteUG is unlearnable because we hear only UG positive examples, no negative examples. Let’s imagine we were able to find some negative examples. What would a negative example look like? Would it stand out as a negative example, or would we not be able to immediately tell it apart from a positive example? If we found proper negative examples, could we show them do a T2/T3 robot and would it be able to learn UG? I am not sure how we would ever program UG into a robot (and I therefore don’t have a clue how it is programmed into our own brains). But that is just the easy problem.
ReplyDelete