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.
Subscribe to:
Post Comments (Atom)
-
Bekoff, M., & Harnad, S. (2015). Doing the Right Thing: An Interview With Stevan Harnad . Psychology Today . Twelve years after steppi...
-
Harnad, S. (2003) The Symbol Grounding Problem . Encylopedia of Cognitive Science. Nature Publishing Group. Macmillan. or: Harnad, S...
-
2b. Harnad, S. (2008) The Annotation Game: On Turing (1950) on Computing, Machinery and IntelligenceHarnad, S. (2008) The Annotation Game: On Turing (1950) on Computing,Machinery and Intelligence . In: Epstein, Robert & Peters, Grace (...
It is clear that Pullum and Scholtz do not distinguish universal grammar (UG) from ordinary grammar (OG).
ReplyDeleteIn response to my comment for the previous reading, I realize that my question should refer to OG rather than UG, because OG can be learned, but UG cannot be learned. Regarding the poverty of the stimulus argument, given that UG errors do not exist, they cannot be demonstrated. In other words, there is no negative evidence for UG because we are only exposed to UG-compliant examples. So, we somehow “know” UG without having to learn it.
However, negative evidence for OG can be demonstrated (mostly by children, but adults can make these errors, too). Even though a child may not be directly corrected for violating OG rules (i.e., through instruction), it is possible that there is still corrective feedback involved. If the child realizes they are not understood or does not get the response they are looking for, I’m sure they would eventually learn to correct their mistake. The child can also learn OG rules by watching others communicate to one another. So, I assume it would be possible for the child to learn OG through instruction as well as supervised and unsupervised learning.
I agree with your comment, but one of the things I cannot place in the OG negative or corrective feedback discourse is that children will make mistakes in their speech and not notice, but they will notice the same mistakes in adult speech. I am just wondering how that contributes to the learning of a language and what kind of learning this is?
DeleteThe second part of your comment reminds me of a critique held by anti-UG psychologists of language to which I have been exposed in a previous seminar at McGill. M. Christiansen and C. Chater were critical of the necessity of a language specific module (i.e., an unlearnable UG). They stated that learning a language is essentially a problem of coordination with others rather than a problem of pure categorization. In other words, learning a language would be to learn to do like others are doing rather than extracting the features of the category of “correct sentences”.
DeleteWhile this line of argument is certainly true for OG, it is not clear if it would work for UG. Indeed, I think that a key strength of UG is that we can tell instinctively if a sentence in our native language is wrong even if we never heard anything like that before. But I’m not sure if that ends the issue. Indeed, I imagine that one could argue that we recognize odd sentence as wrong because they are dissimilar enough from the average sentences that we heard and learned by imitation. But at the same time, we could counter this argument by providing a very long and difficult (but still grammatically correct) sentence that is probably dissimilar to anything we are habituated to interact with, and we would be able to tell if the sentence is correct.
---
As a side note, I am confused about finding a “pure” UG violation example. Is it even possible? For example, in a previous class in linguistics, they gave the example pairs of sentences to expose UG violation:
The judges chose a picture of Tyler. (Correct)
Who did the judges choose a picture of? (Correct)
Vs
A picture of Tyler won first prize. (Correct)
Who did a picture of win first prize? (Incorrect)
But now in PSYC 538, I am not sure if this example would count as an OG example or a UG example (after all, every UG rules has to be instantiated in language a way or another).
I’m not sure I really understand how to intuitively distinguish between OG and UG. For example, it seems clear to me that the plurals in noun-noun compounding discussion is OG. But the ‘anaphoric one’, and maybe even ‘auxiliary sequences’ - it would be my hunch that these are UG, because I can’t imagine someone violating these rules - I can’t even imagine an infant violating these rules in these types of constructions.
DeleteCan someone help me understand why these might actually be examples of OG?
Hi Erin, first my trick to consider how language might be to children, is to try to think about language in my third language instead of in my first. When I think about expressing myself in a language that I'm not super comfortable with, there are some givens about language that start to disappear.
DeleteNow about your question, in the example of the anaphoric one, I might have intuitively marked that as part of UG too, but if I had to think about why it wouldn’t be, I would say something like this. In the text, the sentence is the following: “John has a blue glass but Alice doesn’t have one.” The replacement of “a blue glass” by “one” seems to be one that's made for the sake of eloquence/ avoiding repetition, but the sentence in the form “John has a blue glass but Alice does not have a blue glass.” would still be grammatical, make sense and carry all of the meaning it needs to. Thus, I would say this is probably something we learn and that just makes sentence keep their meaning while making them easier and prettier. Because UG must be something that's common in every possible language, and because it would be easy for me to imagine a language in which that replacement didn't exist, I would argue that the anaphoric one is learned (OG), probably through unsupervised learning. On the contrary, breaking a rule of UG like the subject-predicate structure of a sentence does not allow for language to convey what it’s supposed to. For example, if we said “has a blue glass”, that would not be a sentence and that would not have meaning the same way “John has a blue glass” does, so subject-predicate must be UG. And indeed, I don't know of a language that breaks the subject-predicate structure, even if the subject is sometimes implicit like in Spanish, it's there. I hope that’s helpful.
I don't know UG, so I can't state the rules that make a UG error a UG error. But I do know what OG rules make OG errors OG errors. And I know that those rules are learned, and learnable (by unsupervised and supervised learning, or of course by instruction).
DeleteUnless you have really mastered UG rules (to the point where you can teach them to kid-sib by verbal instruction) I don't think you can explain or get insight into them through your knowledge (implicit or explicit) or intuitions about OG. You may think you can explain a particular instance of a UG violation on the basis of OG, but you cannot explain the infinite other instances of that same UG violation (and you've probably unknowingly taken UG for granted in your explanation).
Also, both OG violations and UG violations sound intuitively "wrong." In the case of the OG violation you may or may not know explicitly which OG rules they violated; with UG violations they just sound wrong, and you really don't know why, even if you think you have an OG explanation. There's no escaping the fact that UG is a technical field and you need the technical knowledge, not just OG rules or intuitions, to understand and explain how and why a UG violation is wrong.
I don't have that technical knowledge, and neither do any of you (even if you have taken undergrad linguistics).
That's the best I can do...
Antoine made an interesting point saying that “after all, every UG rules has to be instantiated in a language one way or another”. This is especially relevant when reflecting on the possible implementation of UG in AI agents.
DeleteAs what was said by Professor Harnad in a previous skywriting, UG is syntax but the syntax of a natural language whose words have meaning (reading 9a). When linguists are speaking of UG, they need to adopt a meta-language in which they are testing strange sentences that are not normally uttered by speakers of the given meta-language-- the latter being French, English, etc.… But all meta-languages that linguists adopt are natural languages in which words have meaning and are grounded. The problem is that computational languages are formal; the symbols used for coding have no intrinsic meaning. So even if linguists were to discover the rules of UG and some engineers were to implement them in a robot, all of what we can do at present is to code these rules in a formal language.
Linguists can make sense of UG rules because they already have a language that uses UG. It is similar to learning a new language. Indeed, when learning a second language, you don’t have to ground words all over again because they are already grounded in your first language. This allows us to simply learn the second language by instructions, written or said, in our first language. We are basically learning a new flavor of the same UG governed set. But formal languages are not governed by UG. So even if we could artificially implement these UG rules in an AI agent, it would not be able to learn any languages without proper grounding.
“But each such success is likely to eliminate or weaken some particular apparent APS instance, by revealing that more can be learned in a data-driven way than was previously thought. To some extent this has already begun to happen. The literature on ìunsu- pervised learningî in computational linguistics (Brent 1993 and Sch ̧tze 1995 are two examples) is demonstrating that algorithms can be learned much more from text than most linguists would have thought possible.”
ReplyDeleteThe authors gesture towards research into data driven, unsupervised learning and claim that those defending the APS will soon need to address this work. This paper came out in 2002, and I'm curious to see if that line of inquiry has seen any traction/breakthroughs since then. If impressive work in this field has come about, there is a serious question of its validity in addressing the APS. A recent and celebrated paper detailing a system that mastered Go via unsupervised learning (Silver et al 2017) claimed that their algorithm reached a ‘superhuman’ level at the game without human knowledge, though it employed and depended on hard coded machinery specific to the domain of Go. A paper attempting unsupervised UG learning with similar problems might only bolster the strength of the APS.
No, it has not since been demonstrated that UG can be learned through unsupervised learning based on the data available to the child. No unsupervised learning algorithms can learn UG from the patterns and correlations in what the child says and hears. Neither can linguists; they need trial and error and hypothesis-testing, guided by the UG they all have in their heads.
DeleteThat the game of Go (or chess, or any other capacity that is not UG) can be learned from data through unsupervised learning is neither here nor there, when it comes to the learnability of UG from the child's data.
And the reason UG is unlearnable continues to be the absence of negative evidence: no mistakes made, corrected, heard, or heard corrected. That is the poverty of the stimulus (APS), it pertains also to supervised learning, and it remains unlearnable even if the language-learning child's database is magically increased to include every sentence ever heard or spoken by anyone (except of course the sentences spoken by teams of Chomskian linguists in discovering and explaining UG rules in the 60 years or so since Chomsky discovered and began studying them).
In the article the authors state that there are two ways children learn language: innate priming or data-driven learning. Innate priming involves inherent language modules already developed in a child’s brain at birth. Data-driven learning involves trial and error feedback and life experiences. Pullum and Scholz outline, what they believe, is the argument from poverty of stimulus (APS) based on these two ways to learn language. They state:
ReplyDelete“a. Human infants learn their first languages either by data-driven learning or by innately-primed learning.
b. If human infants acquire their first languages via data-driven learning, then they can never learn anything for which they lack crucial evidence.
c. But infants do in fact learn things for which they lack crucial evidence.
d. Thus human infants do not learn their first languages by means of data-driven learning.
e. Conclusion: human infants learn their first languages by means of innately-primed learning.”
The authors then argue that the soundness of this argument rests on whether premise c is true, which they believe to be unsubstantiated by current research. I believe that Pullum and Scholz were caught up on the wrong premise. In premise d, I do not believe that APS claims that children do not learn their first language by means of data-driven learning. Rather, I believe APS argues that children do not learn their first language solely by means of data-driven learning.
As we have discussed in class, there is a crucial difference between universal grammar (UG) and ordinary grammar (OG). OG is the rules of grammar contained in a specific language, such as English or French. UG is the rules of grammar for all languages. The OG for any language is thus a subset of UG. Some aspects of language are learned such as OG or vocabulary. But UG cannot be learned, because violations of UG are never seen in any natural language. The argument that UG is innate is not the APS argument at all, because there isn’t just not poverty of stimulus, there is no stimulus at all. Pullum and Scholz attempt to critique the APS that they reconstructed, not the true APS that Chomsky outlines.
I disagree that OG is a subset of UG.
DeleteI think UG are those structures that are true across all languages (subject to the parameters of individual languages), and argued to be innate, but OG is just the specific grammar rules (or symbol-manipulations) that are acceptable in an individual language.
OG isn't contained in UG, in the sense that, the grammar rule that in English, an indirect object appears after the verb - is not contained in UG. I would argue it's the other way around - UG is a subset of OG.
Also, I could be wrong about this, but I think the "poverty of the stimulus" is the argument that there is no stimulus at all for UG non-compliant utterances – or that children are not exposed to enough data to learn all the grammar rules in a given language (most strongly because they don't hear UG non-compliant phrases – so they can’t “learn” UG – it must be innate); in other words, I think it’s called “the poverty of the stimulus” because some stimuli are lacking – ie UG non-compliant utterances.
Yes, the poverty of the stimulus is the absence of UG-errors (negative evidence), and hence also the absence of corrections of these never-made errors.
DeleteSo UG is not learnable by the child through supervised (trial/error/corrective-feedback) learning, which requires both positive and negative evidence.
This fact is completely obscured if a distinction is not made between UG-errors (none) and OG errors (plenty). Pullum does not make this distinction, and freely mixes up the two.
Pullum also freely mixes up unsupervised and supervised learning. If the child's data include no negative evidence (UG-errors), then magically augmenting their data-base by adding every sentence ever spoke or written by anyone would not help, even if it included some accidental UG-violating typos, or even the written and spoken work of teams of linguists discussing UG violations such as "John is easy to please Mary."
Unsupervised learning is an even weak form of learning than supervised learning.
Questions for the whole class:
ReplyDelete-- Now that you you know the difference between UG and OG
-- now that you know the difference between unsupervised learning and supervised learning
-- now that you know that the "poverty of the stimulus" means the absence of negative evidence (in the case of UG, the absence of UG-errors, hence of UG corrections), making it impossible to learn what distinguishes UG-compliant and UG-violating utterances
-- yet the child nevertheless speaks UG-compliantly, makes only OG errors, and OG is learnable:
(1) How and why do you think UG evolved?
(2) What adaptive advantages did UG confer, over and above the adaptive advantages of language itself (universal propositionality?)
(3) Did UG evolve all-at-once or gradually?
(4) If gradually, what were the intermediate stages and their advantages? Were the intermediate stages languages? Were they propositional?
I started to think on your question of " Did UG evolve all-at-once or gradually?" and it got me thinking about grammar in communication outside of syntax. Take miming for example: today when we play a game for charades, we often make movements and gestures that replicate the linguistic grammar of the word or statement we are trying to communicate to the other players. Taking this a step further I can see parallels to this in more abstract ideas.
DeleteTake for example early cave drawings. Those drawings don't have strict verbs and nouns like language we consider today, but looking at some I quickly googled, I can generally understand which pictograms were used to depict an object and which were used to depict an action. Is this not a grammar of sorts in a pre-"language" means of communication? We would establish the things we are communicating about then in the style of the pictogram, communicate the action those beings/things are being used for and their effect.
I have been thinking about your first question: “How and why do you think UG evolved?” In the class, we talked about the problem of the evolution of UG. If there was a language before UG existed, how did that language start without the existence of UG? If there was no language before UG evolved, how could the genes that accounted for UG have been selected since they do not seem to give an advantage in the lack of language?
DeleteIn other words, it doesn’t look possible to explain the evolution of UG in terms of the language. Maybe UG evolved out of the daily routines and the environmental cues, which lead to some learned rules with regards to how the environment worked. For example, you need to find food, only then you can eat the food. Statistically, this is always the way things should happen: first get food, second eat the food. So, this temporal association rule between finding the food, and eating the food might have been a precursor to UG rules. Or a non-temporal cue, like when you see the darker clouds and rain together, you get to put these two things together since, statistically, they appear together more often, they don’t. I do not say that association learning leads to UG. I think that these types of rules with regards to the relation of two (or more things) might have led to the development of UG rules. And the genes related to the ability to learn these types of statistical rules might have been selected since they increased the probability of survivor by leading the individuals to act more efficiently, safely in their environment.
However, I feel like there should be another component to what lead to the UG. It looks like other animals would also benefit from efficient and safe navigation in the environment. Maybe though, the environment became more and more complex for humans which lead to a selection for better learning of these rules.
1) On why UG evolved: If we agree that humans have an innate capacity for UG, and that the aim of evolution is survival and reproduction, and that language allows for higher chance of both these aims, the fact that all individuals have a capacity for UG would lead us to the idea that UG benefits language acquisition. Perhaps having an innate ability to sense whether a particular sentence is correct or not, allowed for quicker language learning. Certain combinations of words could thus be ruled out, and thus it was easier to produce the correct combination.
DeleteWhen thinking about the first question, I think that it is unsolvable at the moment. We know that UG must have been an evolutionary advantage in some way but other than that it is impossible to know any reasons that are beyond speculative. For example, in class I proposed that UG may have evolved because it had a positive impact on language learning (for example it would increase the language learning process). However Prof. Harnad explained that assumptions like these are only "assumptions" and have no way to be proven. Hence, I don't think that it would be possible to explain the adaptive advantages of UG.
DeleteWith this being said, I wonder how we would build the "language functions" into our T3 robot Gabe. Considering what we know right now, we know humans have UG, but we don’t know how or why exactly it has evolved. We also know that most of AI algorithms are based on supervised learning and that OG is predominantly learned by instruction and sometimes supervised learning. Furthermore, we know that second-language learning is likely facilitated by supervised and instructional learning. It doesn't seem possible that we could build UG into the robot because we don't know how it works. A question could be; would supervised learning tasks be enough to teach Gabe to human indistinguishable levels? And even if it did, would that even be useful in understanding cognition because we are "by-passing" UG's role in first-language learning?
If UG did evolve as a chance mutation, then it must have first emerged in one individual. But what is the value of language as a communicative tool when only one person has the ability to communicate? Therefore, the evolutionary advantage of language must have been for something else: assisting thought.
DeleteIn addition, developmental psychologists have shown in many studies that human infants seem to have a range of both primate and species-specific learning mechanisms and abilities that enable the acquisition of language. The emerging consensus is that language acquisition can occur without an innate blueprint for grammar.
So, back to the question at hand: Why UG? The issue becomes increasingly complicated when more research is conducted, as the results of many studies undermine the necessity of UG as an innate linguistic template in humans. UG as a thought-assisting mechanism is the only viable theory that makes sense to me. UG definitely aids in language acquisition, but it is possible that its original evolutionary "purpose" was not to allow for facilitated communication.
Hi Emmanuelle! I agree with you, the question of "why UG?" remains puzzling. I, like you, intuit that UG developing for the purpose of assisting thought (cognition). But when I continue to examine this idea I quickly get stuck. I discussed this in the previous reading, but *why* would UG assist us with our cognition? Categorization is central to cognition -- but UG thinking is not necessary for categorization. Consider all of the organisms that categorize despite their lack of capacity for human language. We can never know another's organisms thought or communication mechanism (we can't know for humans either, but UG is a groundbreaking insight into human communication/thought) but we do know that other organisms categorize. So if UG is not necessary for categorization, what could be necessary for? What is it about human thought that benefitted from UG? It aids in language acquisition, but because UG is *inherent* in language. It's a circular explanation that is impossible to get out of.
DeleteIt makes sense that other organisms don't have the capacity for human language/thought (UG style), for the same reason that octopi don't think like cats. Species have specific biologies which result in different ways of experience the world. But there must be something beneficial about it if we *evolved* to use it. Or, as Chomsky has proposed, UG (or the capacity for language) did not *evolve* like any other trait. So how did it get there?
Response to Akhila:
DeleteSimilar to you, I am also curious at how we could possibly build a T3-passing robot capable of producing language. While I do not think how or why language evolved is necessarily important here for input-output equivalence, our lack of understanding of what the UG rules are and as a result, how it can be formalized via an algorithm is a big mystery. While supervised learning could work if we have established the boundaries / all the rules encompassing UG (just like the academics who taxingly referred to the UG they already possessed to teach themselves some UG rules via trial-and-error), it is clear that we do not have the full picture as in real life. I think knowing the UG rules is a prerequisite and necessary step to building a T3-passing robot. However, if we could theoretically build a T3-passing robot just through supervised learning (assuming we also figured out all the UG rules), I do think it would still be useful to understand cognition. It would be a case of underdetermination but the robot would still be able to do everything we do with language. Whether it realistically reflects the course and nature of language acquisition, I do not think so. However, I am not sure how or whether this consideration is particularly necessary if we are trying to just build 'language functions' into Gabe, the T3 robot.
Like Gili and Emmanuelle, I am more drawn to the idea that UG evolved to assist thought. Along with this, I agree that UG is not necessary for categorization since to categorize the right thing with the right kind of thing, both negative and positive evidence is required. With UG, we never receive negative evidence. While this is completely speculative, I found Onur's point about the potential role of daily routines and environmental cues as the basis for when UG started to develop interesting. Specifically, in relation to the step-by-step temporal nature of some high frequency practices (e.g., get food, then eat food). With these practices being necessary for survival, I wonder whether creating relational links between objects to convey this step-by-step nature (as the basis of UG) could've been evolutionary advantageous by allowing members of a group to better structure their thoughts, cooperate with other members in their group and plan for the future. I wonder (while UG is not necessary to categorize), that maybe the evolutionary role of UG could have been to assist thought by helping to better structure categorizes that we do have in ways that ease how we function and communicate in the world.
DeleteResponding to Akhila: "I don't think that it would be possible to explain the adaptive advantages of UG."
DeleteI think that you're right in positing that explaining the adaptive advantages of UG is quite a hard problem. I do not know if it is possible to find these out.
This problem is specifically hard because we have no way of studying people without UG (and, thus, cannot examine how someone without UG would function in social settings, or in learning tasks, etc.).
So, it appears we are stuck. (And, like Prof. Harnad pointed out, we can speculate all we like, but at the end of the day, we have no reason to believe these speculations.
However, while we are not able to find answers to the adaptive advantages of UG, we *can* ask questions about the adaptive advantage of language. And these questions, at least, appear easier (potentially soluble).
In saying this, I do understand that answers to these kinds of questions do not answer the harder (and more important) question about UG.
We don’t even know for sure that UG evolved, let alone why it did.
DeleteThere’s no evidence to consider UG over and above language itself.
UG can’t have evolved gradually because you can’t have partial language.
Kyle, interesting question, but subject/predicate propositions are not just miming or drawing objects and actions. They are predicating -- stating that something is the case, something that is true or false. (How, for example, would you mime "not", as im "the cat is not on the mat." This is not expressed by miming or drawing a cat that is not on a mat...) And when we do charades today, we are a species with the "propositional attitude" already built into our language biased brains; we subtitle everything with a verbal, propositional narrative, whether we like it or not.
DeleteOnur, even without the chicken/egg puzzle it is hard to explain the origin and evolution of UG because it is so complex yet all-or-none (no partial or intermediate UGs exist or make sense). This is not a problem for the evolution of propositionality itself, if it evolved from intentional communication by pantomime and pointing -- except, of course, if there is a Darwinian reason why simple 1st-order Boolean propositional and predicate grammar (OG) would not have been enough for propositionality (i.e., that we cannot really say and mean "the cat is on the mat" [and hence every other possible proposition] unless we "know" (implicitly) that we cannot say and mean "John was easy to please Mary").
This is not a trivial problem. (It's easy to speculate about how experience might have "shaped" UG, but only if one has no idea what UG actually is.)
Anna, but how did that "innate capacity" evolve?
Akhila, it is not just that evolutionary hypotheses are hard to test, but the nature of UG is that it is complex and does not come in inceremental parts. (To build Gabe, we would have to build in UG.)
Emmanuelle, "The emerging consensus is that language acquisition can occur without an innate blueprint for grammar." This "consensus" is emerging as Chomsky (over 90) concentrates his remaining strength on trying to save the world rather than rebut the ever-emerging "consensuses" that he held at bay while still doing UG: How does it aid "thought" that you cannot say "John is easy to please Mary" (and countless other things like that? UG, the explanation, turns out to be complex and all-or-none. It's relation to thought is not known, especially if "UG" is just used as a place-holder for a bunch of syntactic rules without specifying them.)
Gili & Paul, there is no doubt about the adaptive benefit of having propositionality -- but why does propositionality require UG (if it does)?
Madeleine, UG is a universal feature of language, but it is not learned or learnable by the language-learning, so it must have evolved, but how? and why? It cannot be gradual and it is hard to believe that it was one complex mutation.
Rachel, Spot-on...
In reply to Rachel's comment, I'm just wondering why partial language cannot exist and if it's specifically in reference to a first language? Because, if second languages are fair game, then I'd like to argue that partial language is possible, and that in this context a second language would be verbal communication, and first would be drawing/gesturing/basic sounds communication.
DeleteDerya, it's okay don't worry haha thank you for clarifying that for me!
DeleteUpon reading the 4 questions asked by Professor Harnad, I was not quite sure what the relevance of the second question was. As Rachel mentioned in a comment, there is no reason for us to consider UG above language itself. I do not understand the advantages of entertaining the thought that UG could have brought adaptive advantages above and beyond the adaptive advantages of language itself. UG is innate and applies to all languages. There are specific parameters for each language that implicitly tell you which version of UG applies to your language (example: pronoun dropping in certain languages). I do not believe we should ever evaluate and compare the adaptive advantages of UG and language itself because they are so deeply connected. All humans that possess language, possess UG. UG helps us acquire and intuitively know how to use language correctly.
DeleteI also thought the third question asked was strange. It seemed obvious that UG could not have evolved gradually because partial language cannot exist. To possess language, one must be able to say anything and everything using words. People could not have developed partial language that only allowed them to say a small subset of the infinite possible utterances (correct utterances might I add). Then, after giving it more thought, I understood that those who are trying to come up with a theory for how language or UG evolved might entertain the possibility of UG evolving gradually like other traits. Darwinian theory emphasizes that small and gradual variations in genes over generations, this might lead people to think that UG might have evolved gradually as well… but, if you actually think about UG and how it is innate and unlearnable, it would be HIGHLY unlikely (dare I say impossible) that UG evolved gradually.
Re: “UG can’t have evolved gradually because you can’t have partial language.”
DeleteI disagree with you on this point. I think you could have had a UG that would be deemed partial nowadays, but was complete at the time. My understanding is that language gradually evolved, for example it gained in abstraction, moving from pantomimes to very abstract words such as “freedom.” I would argue that all of language and all the rules that apply to a correct use of language were not developed all at once. UG could have evolved gradually, by evolving for the basic, more primitive version of language we started out with and then included new rules as they appeared. But this are only speculations which are unverifiable. I just don’t think we can write that of so definitely. I also can't really imagine something as critical and complex appearing suddenly.
Overall, I find myself in complete agreement with Pullum and Scholz in this article. From my limited reading in linguistics, concepts of universal grammar, language acquisition device and APS, are used with varying interpretations. The authors here provide a clear empiricist description of APS, and evaluate the 4 main studies that support this argument, finding them to be lacking. The sheer infinite possibility of sentences that can be constructed in a certain language abiding to its syntax does not necessarily imply that the ability to do so is innate. As the authors point out, a typical child is exposed to 10 to 30 million word-tokens, thus having access to a high amount of information to form these infinite sentences. Personally, I do not understand why language alone, among other cognitive faculties of memory, categorisation, decision making and others, have been singled out to have been the result of an innate ability.
ReplyDeleteSpecifically, I agree with the authors on negative feedback (page 13). Most of our psychological understanding of behaviour is premised on instrumental learning or reinforcement learning. And often times people learn via exposure to certain stimulus response associations multiple times. This learning also occurs in the absence of any negative feedback. For instance when a child learns to categorise chairs, it does not receive instruction or feedback on all the items that are not chairs. As such the argument for learning language without negative feedback does not seem to be based on premises on which other cognitive phenomena are understood. Or in other words, perhaps people do not make the mistake of adding that to sentences simply because the probably of them having heard such a sentence was low. As such it sounds unfamiliar and possibly wrong.
The replies to your questions are contained in the replies to other postings above, and the last 2 weeks' of skywriting.
DeleteOn re-reading the skywritings, I agree that as a non-linguist I do not understand UG, or know the various sentences that are examples of UG. As such, I do not know if UG exists or not, and understand that most linguists agree that it does exist. Research also shows that parents provide no or little feedback (positive or negative) to children during the language learning process. I see that in “Language Acquisition” Pinker argues against the idea that familiarity can explain children learning a language as they still do not have the rules on how to parse a sentence. So, at best it is a huge dataset from which they can learn implicitly or explicitly.
DeleteRe: Professor Harnad's thought question, did UG evolve all-at-once or gradually?
ReplyDeleteThis inquiry made me curious about what theories had already been proposed. Researchers from the Max Plank Institute for the Science of Human History studied 81 Austronesian languages, examining how grammatical structures and lexicons evolve over time. Their research concludes: once a language is established, grammar changes much more quickly than lexicons, but when a new language is formed, vocabulary changes more quickly. This finding shocked professors of linguistics, as they believed it could challenge the innateness of UG. If we were to entertain this thought, perhaps the dynamic nature of grammar within the studied region points toward a gradually and continually evolving UG.
However, I believe that their findings apply only to OG and not UG. Recognizing that we cannot explicitly know the rules of UG, the researchers' extensive study of grammatical structures belonging to distinct languages seems much more characteristic of OG.
Further, Pallum et al. assert that children "acquire systems that display unexplained universal similarities that link all human languages." The issue with the aforementioned research is that they only studied languages within the constrained region of Austronesia and thus, potentially neglected the property of universality. Taking this into consideration, if UG did actually evolve gradually, would all languages have had to evolve together and at approximately the same rate?
I think once UG was in place, languages could be born much more quickly, since the groundwork for its rules and therefore learning it would already be in place. For the first existing languages, assuming many came be at around the same time, then I would agree they would have to go slowly as UG gradually did. But, again, after UG was in place, languages could evolve rather quickly and be picked up by the next generation.
DeleteTo answer Prof. Harnad's question in a different skywriting, whether UG evolved gradually or all-at-once, I'm not sure we can know. If it was all-at-once, I would imagine the brain structure would have to be in place, and some environmental change either brought out the UG or epigenetically influenced a large majority of the population to favor UG. With a quick google search, I found that Neanderthals had the proper anatomy to be able to form speech, but perhaps Homo sapiens used the presence of Neaderthals as the environmental impetus to use UG. This perhaps enabled better food-gathering and survival, over that of any Neanderthals in the area.
On the other hand, if it were gradual, perhaps there were vocalizations made with common gestures, and eventually everyone in a group made those same vocalizations with the gestures. Gradually, there were rules to which gestures made sense given a context, which formed the early rules. And through time these rules became stronger and the gestures eventually fell out of use but the vocalizations stayed. And then, somehow, it became language. I know you aren't a fan of the "make noise and wow we can communicate!" ideas like those in Sapiens, and I agree. Perhaps though it was gestural, then paired with noises, then they realized you didn't need to see the person you were communicating with in order to get a message across with sound. Thus, vocalizations.
This paper reminds me of a question I had for the argument of the Language Acquisition Device (LAD) of Chomsky and APS. So Chomsky thinks that the babies could learn language easily because of the LAD, and the poverty of stimulus account for the strongest evidence for that. Also, the LAD is no longer there after a certain age, which explains why the second language learning is so hard. Then, I wonder, what in particular is lost in LAD that is lost when people grow up? Because if APS is an evidence of the LAD, then APS is also present in second language learning. If we observe that babies make grammatical mistakes, we should also note that second language speakers do that as well, in spite of locking the negative stimulus of it. Does this fact, by the same logic, manifests that second-language speaker have the evidence for LAD? But if second-language speaker, after certain age, still have LAD, then what explains the extra effort of learning second language compared to the mother tongue?
ReplyDelete
DeleteHi Mingjuan!
That’s an interesting question. I think you can think of the Language Acquisition Device (LAD) as a preparedness to learn in the form of pathways in the brain that are ready to be solidified. That’s why children who haven’t learned language yet are able to recognize many phonemes/morphenes, but aren’t able to do so for all of them after acquiring a language (they know the ones for their language).
In that sense, I don’t think your conception of the LAD is the right one. It’s a preparedness, not an actual structural apparatus in the brain that is sustained throughout the lifetime; hence, critical periods for the acquisition of language. To me, the Argument from the Poverty of the Stimulus is just a way to explain why UG must exists in the critical period, where its parameters are fitted to the language that a person is acquiring.
Mingjuan, I don't think that LAD is completely lost after the critical period for language learning. I think that learning a second language is harder because our brains have already adjusted to the grammar of one language, and learning another requires differential categorization. If I remember correctly, in other classes children's brains are described as being very flexible because connections in the brain haven't been wired yet in the language regions. If this is correct, then it can be used to support an argument for the LAD and UG.
DeleteWhy UG? I have problems with answering this question because I don’t know what language would be without UG and if UG is necessary at all. I like the idea that it could be used to assist with thought, but then again, why is UG there for thought and how does it work? It seems like UG would definitely help with language acquisition, but until I know what language looks like without UG, I would never have an idea of how it evolved in the first place or what it does in its entirety. Could it be that there was once a population with UG and non UG and the fact that having no UG was so debilitating that there was just a complete extinction and wipe out of non UG people?
ReplyDeleteI am unsure as to whether UG evolved gradually or all-at-once. Due to the universality of it and no cases of people without UG, it becomes hard to imagine that it could’ve evolved gradually like many other evolutionary traits. I don’t know if there could have been the same pressure from evolution that we see shaped many of the traits that we now know. I know that UG was discovered because of language, but if we think of UG and language as two different things, we can see a clearer timeline for language as being gradual than UG being gradual. Maybe we all have UG in our brain as organisms, but something in our environment pressured us to use it. And because of how serious that pressure was, it was an all-at-once phenomenon seen in humans. And if we are to say that UG comes as some package deal with some area of our brain, maybe this part of the brain is so primal an important that we can never see a case where someone has no UG because they wouldn’t be alive if they didn’t have it.
I share your confusion around the role of UG and why/how it evolved in the first place. The only way to test the contribution of UG to language acquisition would be to find individuals who lack it and compare them to those who have UG, but that is impossible, since UG is innate. Still, I also think that UG must somehow play a crucial role in language acquisition, but what role exactly (e.g. learning language faster, ‘better’ and/or helping us distinguish between different languages)?
DeleteGiven that UG is innate and specific to the human species, it must be encoded in our genome. The question then is how and why that/these gene/s evolved. If UG were to have evolved gradually, I have difficulty imagining how a ‘partial’ UG would have manifested itself. At what stage would there be ‘enough’ UG for language ability to arise? For an all-at-once scenario, I have trouble thinking of a strong enough motivator for the switch from no UG to complete presence of UG.
I also feel the same with regards to UG. But I often wonder is other animals also have "UG" or the part of our brain that is responsible for UG but just don't have the cognitive capacity to make use of it. As Amy said, it's possible that UG stems from a more primal part of our brain and that maybe environmental pressure caused it to be used. I think it makes sense that as our cognitive skills developed to be more sophisticated, then it's possible that only under those conditions were we able to make use of our UG capacity. I guess I see UG as being similar as our innate ability to categorize, and in this example, cognition would be to UG as what sensorimotor systems would be to categorization
DeletePoverty of stimulus and UG
ReplyDelete“The absence of negative evidence makes it impossible to learn what distinguishes UG-compliant and UG-violating utterances”
I do not disagree with that negative evidence is important when learning a category. But, as has been mentioned in “Uncomplemented Categories, or, What Is It Like To Be a Bachelor”, “all of these uncomplementable categories are experiential categories. And since their complement is nonexperiential, it's either nonexistent or self-contradictory.” It occurs to me that UG is like a (like-)uncomplementable category for people (who are not linguists that look for UG and try to “break the rules”). It even happens on an OG level and people are more aware of OG than of UG – people may report some sentences to be ungrammatically just because they have never heard (/experienced) such saying(/sentence structure) before. “I may not know what exactly it feels like to say a UG-violating thing, but I know what it feels like to say a (UG-compliant) thing because everything I say in a normal context is UG-compliant, and what I am feeling right now is different from that feeling, so I am going to say it is not UG-compliant – maybe it actually is, I don’t know, but it feels different.” Then, is it necessary for one to know “what cannot be spoken” in order to speak “correctly” (or, UG-compliantly, on a UG-syntax level)? Can pure imitation of what has been observed to be working (meaning conveyed) (on a syntax level) be sufficient? (Necessarily innate?) Also, I think if there is UG, for every UG rule, there might be different ways to violate the rules – there is usually one way to do the right thing because there is one right thing, but there can be many different ways to do the wrong thing. In order to capture all the features of UG, all the features have to be violated in some way – would it not be easier if one just takes whatever works and use it?
Also, I am not sure if it is right but can I say “what has been assumed is every OG-compliant utterance is a UG-compliant utterance, and if UG is violated, OG must have been violated”? Then, by integrating everything that OG allows, we are going to get what UG allows. Then, test if the rule we are getting really is a UG rule, we need to see if violating this rule would lead to the violation of every OG.
I'm not sure I understand. UG violations just sound wrong. It feels wrong to say "John is easy to please Mary." Nonlinguists don't know why. There is an infinity of possible UG-compliant (and UG-violating) sentences. The reason one feels right and the other feels wrong is because we "have" the invariants of the category: the rules of UG in our heads already. I don't know about UG connections with OG except that OG rules are in our heads because we learned them. (But maybe I have not understood your questions.)
DeleteThanks for the reply! I think I will rephrase it to make it clearer. (These thoughts are in nature immature yet.) My question is exactly targeting at the feeling of “it just sounds wrong”. Just like you’ve mentioned, it sounds wrong (or right) because “we ‘have’ the invariants of the category: the rules of UG in our heads already”. “Nonlinguists don’t know why”, and linguists have theories of why. Let’s first assume that the theories are all correct. However, are the invariants of the category necessarily innate? It has been said that “the absence of negative evidence makes it impossible to learn what distinguishes UG-compliant and UG-violating utterances”, thus it must have been innate. Yet, with only positive evidences, we are already forming our inferences of “the whole picture”. For example, “layleks”: during the lecture, when you were giving examples of things that are “layleks”, we are already guessing (Bayesian-inferencing) what “layleks” are and are not by finding the commonality of the things that are “layleks”. With the examples that were given, we can “guess” that “layleks” are possibly “entities”, using Nathan’s word, or everything that exists -- it is a reasonable guess, and it could have been right, except it is not (because you said things that do not exist are layleks, too), but it would not have stopped me from correctly listing things that are layleks. Now, back to language. With all the positive evidence of “UG-compliant” sentences, we may already be able to speak “UG-compliantly”. And that is all we need for the purpose of communication – we don’t need to know how is it to speak “UG-violatingly”. Thus, maybe the POS is not a real obstacle in the end, and thus the rules may not necessarily be innate.
DeleteI still agree that without negative evidence, one cannot learn the category; what I am saying is that it seems that one does not need to fully learn the category to use it.
DeleteUG is not just an (infinite) set of possible input strings. It is also an (infinite) set of possible output strings. "Laylek" is just input, and since every possible input string is a member of the category Laylek, whatever hunches you might have about the category's invariants, they are all wrong; and, besides, since Laylek input is all passive (you do not produce layleks), your hunches are all untestable.
DeleteIn contrast, with UG, the child is not just hearing only UG-compliant input but also producing only UG-compliant output, all the time. There is no hypothesis, like "entity," and no hypothesis-testing; just error-free UG-compliant input and error-free UG-compliant output, from the outset; only OG errors and corrections occur, and they are irrelevant to UG (except for parameter-setting).
Early-on, I had a hunch that maybe the way it works is that we hear exclusively UG-compliant input, but when we hear something said that we might have wanted to say, but feel we would have tried to say the same thing differently, non-UG-compliantly ("He just said P, but the way I would have said it is Q, so I guess Q must be wrong"). So in this we are giving ourselves internal error-correcting feedback whenever someone else says something that we feel would have said otherwise. Something like this definitely happens with OG and OG errors. But with OG there are also lots of actual output errors, and external corrective feedback. It would be surprising, to say the least, if UG errors and corrections were all internal and mental only...
I agree that UG is not just an (infinite) set of possible input strings. It is also an (infinite) set of possible output strings. However, I am not sure if I see why "'Laylek' is just input, and since every possible input string is a member of the category Laylek, whatever hunches you might have about the category's invariants, they are all wrong." I am not sure if I see the causal relationship here. Why would the hunches about the invariants *all wrong*? It can be incomplete, which is a type of “wrong”. But just with the incomplete variants, we could already put post-learning things into the category of Laylek successfully (and based on the true invariants of Laylek, we could never find anything that is not a Laylek). And with more learning of the category Laylek, we add more features to the category Laylek, and the invariants approaches to the real invariants. Also, it is true that laylek input is passive, but whenever I ask “is X a laylek?”, I am testing out my hunches. Yes, I may not have *produced* X, but in this case, it seems that X is the output of my categorization.
DeleteI think I partially agree with you on the internal error-correcting feedback of UG, I just do not think it is *that* conscious a process… I think I have come to realize that the using of language and acquisition of language are not entirely *that* conscious – it can be very conscious, but sometimes, no. (Like the participant I mentioned who said “apple was not in the list” and later be so surprised that he said that in stead of “pomme was not in the list”.) Thus, this “error-correcting” might as well be statistical detection (?), which is also something arguably “not on the top of our mind”.
DeleteIt does not work with Laylek because all hypotheses are wrong: there are no non-Layleks.
DeleteYou are right that much (probably most) learning is "unconscious" (i.e., unfelt). That's what makes cogsci difficult: it would be easier if we knew how we learned because we could feel it!
Trying to "brainstorm" the questions asked above:
ReplyDelete1) How and why UG evolved: we don't know, because we cannot find the answer to 2) what adaptive advantages UG confers, over and above the adaptive advantages of language itself (universal propositionality? Since we cannot get any negative data (language violating UG) to analyze its advantages versus disadvantages, separated from those of language itself. It's like a coin of only one side, we can never judge whether the single side is a head or tail.
3) Did UG evolve all-at-once or gradually? I also don't know, but I figure given that there's no negative data (again), it's most likely that it cannot evolve gradually. So there's no room left for question 4) If gradually, what were the intermediate stages and their advantages? Were the intermediate stages languages? Were they propositional?
"*How* a grammar is constructed is not the same question as *why* that particular grammar was constructed" (2002). This quote really hammered in what we have been discussing within the scope of evolution and language as well as in this course as a whole. I think this is a very important point to be made when talking about UG because it highlights that figuring out how UG works/how UG was created or came about is not the same as figuring out *why* it came about. If we learn all the rules of UG for example and really grasp how it works, this will not tell us why it emerged in the first place. This to me is very reminiscent of Fodor's argument regarding brain activity, neural imaging, etc.. Fodor's claim was that figuring out where and when something happens in the brain is not the same as figuring out how and why it happened. This resembles the point about UG to me because it is stating that figuring out how (the rules, parameters, restrictions, etc.) UG works is not the same as figuring out why we have UG. Why did UG evolve? It seems that we can create two questions generally resembling the easy and hard problems, regarding UG: How does UG work? And why did UG evolve? Similarly to the real easy and hard problem, we are still so far from solving the easy problem - we are also still quite far from figuring out how UG works at all. This is due to many factors, one of the most powerful being that UG is uncomplimented. Meaning that we never hear non-compliant UG - this helps us to solve that UG is in fact innate. However, it makes it very hard to solve how it evolved or - even more difficult - why it evolved. As discussed in class, there is no such thing as half UG which makes it very difficult to distinguish the rules of UG and furthermore, how/why it evolved.
ReplyDeleteYes! Finding how and why UG evolved belongs to the same category of questions about why and how we feel, for which we have no answer for! And like the hard problem being in my opinion insoluble, knowing why and how we have UG may also not be soluble and it is certainly not evolutionary accounts that will help us solve this question. This course has led me to believe that many of the questions that we have about life will remain just that, questions, with no apparent answers.
DeleteIn relation to the articles read during week 8 when trying to understand why did language evolve with humans but not with other species , one hypothesis was that humans vocalized propositions because they were motivated to do so and that there may have been a genetic mutation that conferred them an adaptive advantage through natural selection. Could the same be said with UG? That having UG conferred an adaptive advantage for learning language and thus surviving. Or that it may have been through a process of motivation? Once UG was genetically coded, a certain part of the brain , or certain networks would only be dedicated to UG, similarly to have certain neural networks dedicated to code memory in the hippocampus or recognize objects in the temporal lobe for example.
ReplyDeleteBut say you had a mutation for UG - how would that be beneficial if no one else had it? UG wouldn't be adaptive to survival if it doesn't aid in communicating with others who did not have UG.
DeleteAnd what do you mean by "it may have been through a process of motivation?" How could one's motivation lead to them having UG?
I think that it's pretty reasonable to infer that there was an adaptive advantage to having UG if all humans have it now. We just don't know the exact distal cause of that, and we don't have any negative evidence to compare it with.
DeleteLike Emily, I'm also confused about what you mean by "a process of motivation". Is this more of a speculation about whether we somehow had UG coded capacity and weren't using it until we were motivated to do so, or if being motivated to make propositions lead to UG being encoded?
Reading over the papers this semester and everyone’s responses, I have come to question the validity of UG theory. I may very well be lost in the sauce… How are we able to say UG theory, a set of innate grammatical rules, is false? Answer: if a child makes a UG mistake and someone corrected her. Yet, we don’t have negative evidence. Does this make UG theory unfalsifiable? I could very well be mistaken… did I incorrectly treat the definition of UG as the theory of UG? I am a bit confused.
ReplyDeleteHey William, can certainly empathize with being 'lost in the sauce', especially regarding the linguistic portion of our class which has had my head spinning from the beginning. Like you've said yourself, the poverty of the stimulus of Chomsky's Universal Grammar necessarily ensures no negative evidence, which would be required to ensure its existence. AS for your second question, I am also pretty unsure, but what I can say is that like many others in the comment thread, the lack of any negative evidence is analogous to the hard problem. Professor Harnad brought this up in a later class: yes we can't answer why something feels like it does, but we also can't answer what it feels like to not feel anything. The circular nature of the argumentation might actually prove to be its downfall, as without any real ability to test the absence of UG (through corrective learning) we can't confirm any of our hypotheses regarding the how question surrounding it. Furthermore, what Chomsky defines as UG could maybe be understood in a completely different way that is unrelated to its syntaxical explanation in the vein of weak equivalence? Who knows?
DeleteHi William,
DeleteI share the same question as you that I don’t understand how UG can be proven wrong if it’s innate. To me UG seems very plausible. Humans have a strong motivation to learn language, however only being motivated is not enough, I believe that there is a mechanism that helps us through this motivation, and it is UG. Also, studies done with pregnant women in the thirst trimester prove that the baby can differentiate one language from another. This to me is a huge proof that some portion of language is innate and is not learnable.
Stimulus poverty is supported by the example given in my comment on reading 9a: if 2 children of the same age (toddlers for the sake of this example), one with parents with highschool diplomas and the other with parents with PHd’s are compared, there is a high chance that the later will have a vocabulary almost double of the other. In this particular study the conclusion is drawn that this is due to being exposed to a richer vocabulary in their every day environments.
ReplyDeleteA point must also be made to the type of stimulus – children cannot learn language (and show the progress in the long run) without real life human interaction. Studies have been conducted using books, audio recordings, even tv recordings, however without the interaction and active feedback provided by a real person being present to teach a language, the child does not effectively learn the new vocabulary.
The authors appear to believe that APS depends on positive evidence of OG. According to them, ASP is a problem resulting from parents not using sophisticated language when speaking with kids, which prevents kids from learning the rules of grammar. However, ASP doesn’t come about because of baby talk. What is of interest is how babies learn to speak regardless of never hearing or speaking in non-UG-compliant ways and also never being corrected (for their non-existent UG mistakes) and never receiving instruction. And yet the authors err by dismissing the necessity of negative evidence as an exception that is irrelevant to APS – because negative evidence is in fact very important for category learning (the Laylek example clearly indicates this). They say that: “For reasons of space, we do not attempt to give further details here. There is of course much more to be said about whether negative data are crucially necessary for learning natural languages from it.” This would suggest that the authors miss a key component of learning and what APS is all about.
ReplyDelete