Thursday, April 30, 2015

Moocs being assessed

Those interested in this topic might find the following interesting (here). It is now the received wisdom that MOOCs were hyped when started and their promise, if there is any, was largely in the minds of those that stood to benefit most. The Gates foundation is now funding research on this topic (which, given the monetary source may or may not be "objective," pardon my cynicism) so that we can find out what MOOCs are good for, if anything.

It appears that the one group that they do service are "non-traditional" students and the problem appears to be keeping their attention.

The report names student engagement as a prominent theme. Many students enrolled in MOOCs are nontraditional, so making sure that they are engaged and able to succeed in such a course is even more important. Figuring out how to maintain students’ interest during an online course when “a distraction is literally just a click away” is another important element, Mr. Siemens said.

So putting things on line has some potential drawbacks that researchers are now addressing. Note too the audience, "non-traditional" students. It seems that for the regular college crowd MOOCs may not be on the agenda. Effectively, MOOCs are now filling the role that correspondence courses filled in the pre-digital era. And it seems that they are finding problems analogous to those that such courses traditionally encounter; keeping the student's attention focused on the material. This does not strike me as very surprising. It was never clear to me why presenting the material on line on a screen should make it more engaging than doing so in a book on your lap. At any rate, the discussion goes on, this time with much less hype.

Wednesday, April 29, 2015

Shigeru Miyagawa Vitor Nóbrega comment on the previous post

Thanks to Shigeru and Vitor for taking the time to elaborate on the points they make in their paper.


Dear Norbert,

Thanks for taking up our paper in your blog (Nóbrega and Miyagawa, 2015, Frontiers in Psychology). We are glad that you appreciate our arguments against the gradualist approach to language evolution. There are two things that don't come out in your blog that we want to note.

First, our arguments against the gradualist view are predicted by the Integration Hypothesis, which Miyagawa proposed with colleagues in earlier Frontiers articles (Miyagawa et al. 2013, 2014). The gradualists such as Progovac and Jackendoff claim that compounds such as doghouse and daredevil are living fossils of an earlier stage in language, which they call protolanguage. The reason is that the two "words" are combined without structure, due to the fact that these compounds (i) have varied semantic interpretations (NN compounds), and (ii) are unproductive and not recursive (VN compounds). We argued that if one looks beyond these few examples, we find plenty of similar compounds that are fully productive and recursive, such as those in Romance and Bantu. These productive forms show that the members that make up the compound are not bare roots, but are "words" in the sense that they are associated with grammatical features of category and sometimes even case.

This is precisely what the Integration Hypothesis (IH) predicts. IH proposes that the structure found in modern language arose from the integration of two pre-adapted systems. One is the Lexical system, found in monkeys, for example. The defining characteristic of the L-system is that it is composed of isolated symbols, verbal or gestural, that have some reference in the real world. The symbols do not combine. The other is the Expressive system found in birdsong. The E-system is a series of well-defined, finite state song patterns, each song without specific meaning. For instance, the nightingale may sing up to 200 different songs to express a limited range of intentions such as the desire to mate. The E-system is akin to human language grammatical features. These are the two major systems found in nature that underlie communication. IH proposes that these two systems integrated uniquely in humans to give rise to human language.

Based on the nature of these two systems, IH predicts that the members of the L-system do not combine directly, since that is a defining characteristic of the L-system. E must mediate any such combination. This is why the IH predicts that there can't be compounds of the form L-L, but instead, IH predicts L-E-L. Such an assumption bears a close relation to how human language roots are ontologically defined, as feature-less syntactic objects. Once roots are feature-less they are invisible to the generative system, thus there is no motivation a priori to assume that syntax merges two bare roots, that is, two syntactically invisible objects. 

The second point is that the L-system is related to such verbal behavior as the alarm calls of Vervet monkeys. We focus on the fact that these calls are isolated symbols, each with reference to something in the real world (thus, they are closer to concepts rather than to full-blown propositions). You question the correlation by noting that while the elements in a monkey's alarm calls appear purely to be referential, words in human language are more complex, a point also Chomsky makes. We also accept this difference, but separate from this, roots and alarm calls share the property, if we are right, that they are isolated elements that do not directly combine. This is the property we key in on in drawing a correlation between roots and alarm calls as belonging to the L-system. In addition to the referential aspect of alarm calls, there is another important question to solve: what paved the way to the emergence of the open-vocabulary stored in our long-term memory, since alarm calls are very restricted? Perhaps what you’ve mentioned as “something ‘special’ about lexicalization”, that is, the effect that Merge had on the pre-existing L-system, may have played a role in the characterization of human language roots, allowing the proliferation of a great number of roots in modern language. Nevertheless, we will only get a satisfactory answer to this question when we have a better understanding of the nature of human language roots.

Finally, you might be interested to know that Nature just put up a program on primate communication and human language on Nature Podcast in which Chomsky and Miyagawa are the linguists interviewed.
Also, BBC will be broadcasting a Radio 4 program on May 11 (GBST) about evolution of language that will in part take up the Integration Hypothesis (or so, Miyagawa was told).

Shigeru Miyagawa
Vitor Nóbrega

April 29, 2015

Tuesday, April 28, 2015

A minimalist evolang story

MIT News (here) mentions a paper that recently appeared in Frontiers in Psychology (here) by Vitor Norbrega and Shigeru Miyagawa (N&M). The paper is an Evolang effort that argues for a rapid (rather than a gradual) emergence of FL. The blessed event was “triggered” by the emergence of Merge which allowed for the “integration” of two “pre-adapted systems,” one relating to outward expression (think AP) and one related to referential meaning (think CI). N&M calls the first the E-system and the second the L-system. The main point of the paper is that the L-system does not correspond to anything like a word. Why? Because words found in Gs are themselves hierarchically structured objects, with structures very like the kind we find in phrases (a DMish perspective). The paper is interesting and worth looking at, though I have more than a few quibbles with some of the central claims. Here are some comments.

N&M has two aims: first to rebut gradualist claims concerning the evolution of FL. The second is to provide a story for the rapid emergence of the faculty. I personally found the criticisms more compelling than the positive proposal. Here’s why.

The idea that FL emerged gradually generally rests on the idea that FL builds on more primitive systems that went from 1-word to 2-word to arbitrarily large n-word sequences.  My problem with these kinds of stories has always been how we get from 2 to arbitrarily large n. As Chomsky has noted, “go on indefinitely” does not obviously arise from “go to some fixed n.” The recursive trick that Merge embodies does not conceptually require priming by finite instances to get it going. Why? Because there is no valid inference from “I can do X once, twice” to “I can do X indefinitely many times.” True, to get to ‘indefinitely many X’ might casually (if not conceptually) require seguing via finite instances of X, but if it does, nobody has explained how it does.[1] Brute facts causing other brute facts does not an explanation make.

Let me put this another way: Perhaps as a matter of historical fact our ancestors did go through a protolanguage to get to FL. However, it has never been explained how going through such a stage was/is required to get to the recursive FL of the kind we have. The gradualist idea seems to be that first we tried 1-word sequences then 2-word and that this prompted the idea to go to 3, 4, n-word sequences for arbitrary n. How exactly this is supposed to have happened absent already having the idea that “going on indefinitely” was ok has never been explained (at least to me). As this is taken to be a defining characteristic of FL, failing to show the link between the finite stages and the unbounded one (a link that I believe is conceptually impossible to show, btw) leaves the causal relevance of the earlier finite stages (should they even exist) entirely opaque (if not worse).  So, the argument that recursion “gradually” emerged is not merely wrong, IMO, it is barely coherent, at least if one’s interest is in explaining how unbounded hierarchical recursion arose in the species.[2]

N&M hints at a second account that, IMO, is not as conceptually handicapped as the one above. Here it is: One might imagine a system in place in our ancestors capable of generating arbitrarily big “flat” structures. Such structures would be different from our FL in not being hierarchical, and the same in being unbounded. These procedures, then, could generate arbitrarily “long” structures (i.e. the flat structures could be indefinitely long (think beads on a string) but have 0-depth).  Now we can ask a question: how can one get from the generative procedures that deliver arbitrarily long strings to our generative procedures which deliver structures that are both long and deep? I confess to having been very attracted to this conception of Darwin’s Problem (DP). DP so understood asks for the secret sauce required to go from “flat” n-membered sets (or sequences for arbitrary n) to the kind of arbitrarily deeply hierarchically structured sets (or graphs or whatever) we find in Gs produced by FL. I have a dog in this fight (see here), though I am not that wedded to the answer I gave (in terms of labeling being the novelty that precipitated change). This version of the problem finesses the question of where recursion came from (after all, it assumes that we have a procedure to generate arbitrarily long flat structures) and substitutes the question where did hierarchical recursion come from. At any rate, the two strike me as different, the second not suffering from the conceptual hurdle besetting the first.

N&M provides more detailed arguments against several current proposals for a gradualist conception for the evolution of FL. Many of these seem to take words as fossils of the earlier evolutionary stages. N&M argues that words cannot be the missing link that gradualists have hoped for. The discussion is squarely based on Distributed Morphology reasoning and observations. I found the points N&M makes very much to the point. However, given the technical requirements needed to follow the details, I fear that tyros (i.e. the natural readership of Frontiers) will remain unconvinced. This said, the points seem dead on target.

This brings us to the second aim of the paper, and here I confess to having a hard time following the logic. The idea seems to be that Merge when added to the E systems we find in bird song and the L system we find in vervets gets us the kinds of generative systems we find in G products of FL  This is a version of the classical Minimalist answer to DP favored by Chomsky. I say “sort of” as Chomsky, at least lately, has been making a big deal of the claim that the mapping to E systems is a late accretion and the real action is in the mapping to thought. I am not sure that N&M disagrees with this (the paper doesn’t really discuss this point) as I am not sure how the L-system and Chomsky’s CI interface relate to one another. The L-system seems closer to concepts than full-blown propositional representations, but I could be wrong here.  At any rate, this seems to be the N&M view.

Here’s my problem; in fact a few. First, this seems to ignore the various observations that whatever our L-atoms are they seem different in kind from what we find in animal communication systems. The fact seems to be that vervet calls are far more “referential” than human “words” are. Ours are pretty loosely tied to whatever humans may use words to refer to. Chomsky has discussed these differences at length (see here for a recent critique of “referentialism”) and if he is in any way correct it suggests that vervet calls are not a very good proxy for what our linguistic atoms do as the two have very different properties. N&M might agree with this, distinguishing roots from words and saying that our words have the Chomsky properties but our concepts are vervetish. But how turning roots into words manages this remains, so far as I can see, a mystery. Chomsky notes that the question of where the properties of our lexical items comes from is at present completely mysterious. But the bottom line, as Chomsky sees it (and I agree with him here), is that “[t]he minimal meaning-bearing elements of human languages – word-like, but not words -- are radically different from anything known in animal communication systems.” And if this is right, then it is not clear to me that Merge alone is sufficient to explain what our language manages to do, at least on the lexical side. There is something “special” about lexicalization that we really don’t yet understand and it does not seem to be reducible to Merge and it does not seem to really resemble the kinds of animal calls that N&M invokes. In sum, if Merge is the secret sauce, then it did more than link to a pre-existing L-system of the kind we find in vervet calls. It radically changed their basic character. How Merge might have done this is a mystery (at least to me (and, I believe, Chomsky)).

Again, N&M might agree, for the story it tells does not rely exclusively on Merge to bridge the gap. The other ingredient involves checking grammatical features. By “grammatical” I mean that these features are not reducible to the features of the E or L systems. Merge’s main grammatical contribution is to allow these grammatical features to talk to one another (to allow valuation to apply). As roots don’t have such features, merging roots would not deliver the kinds of structures that our Gs do as roots do not have the wherewithal to deliver “combinatorial systems.” So it seems that in addition to Merge, we need grammatical features to deliver what we have.

The obvious question is where these syntactic features come from?  More pointedly, Merge for N&M seems to be combinatorically idle absent these features. So Merge as such is not sufficient to explain Gish generative procedures. Thus, the real secret sauce is not Merge but these features and the valuation procedures that they underwrite. If this is correct, the deep Evolang question concerns the genesis of these features, not the operation instructing how to put grammatical objects together given their feature structures. Or, put another way: once you have the features how to put them together seems pretty straightforward: put them together as the features instruct (think combinatorial grammar here or type theory). Darwin’s Problem on this conception reduces to explaining how these syntactic features got a mental toehold. Merge plays a secondary role, or so it seems to me.

To be honest, the above problem is a problem for every Minimalist story addressing DP. The Gs we are playing with in most contemporary work have two separate interacting components: (i) Merge serves to build hierarchy, (ii) AGREE in Probe-Goal configurations check/value features. AGREE operations, to my knowledge, are not generally reducible to Merge (in particular I-merge). Indeed trying to unify them, as in Chomsky’s early minimalist musings, has (IMO, sadly) fallen out of fashion.[3] But if they are not unified and most/many non-local dependencies are the province of AGREE rather than I-merge, then Merge alone is not sufficient to explain the emergence of Gs with the characteristic dependencies ours embody. We also need a story about the etiology of the long distance AGREE operation and a story about the genesis of the syntactic features they truck in.[4] To date, I know of no story addressing this, not even very speculative ones. We could really use some good ideas here (or, as in note 3, begin to rethink the centrality of Probe/Goal Agree).

I don’t want to come off sounding overly negative. N&M, unlike many evolangers know a lot about FL. Their critique of gradualist stories seems to be very well aimed. However, precisely because the authors know so much about FL while trying to give a responsible positive outline of an answer to DP the problem, the paper makes clear the outstanding problems that providing an adequate explanation sketch faces. For this alone, N&M is worth reading.

So what’s the takeaway message here? I think we know what a solution to DP in the domain of language should involve. It should provide an account of how the generative procedures responsible for the G properties we have discovered over the last 60 years arose in the species. The standard Minimalist answer has been to focus on Merge and argue that adding it the capacities of our non-linguistic ancestors suffices to give them our kinds of grammatical powers. Now, there is no doubting that Merge does work wonders. However, if current theoretical thinking is on the right track, then Merge alone is insufficient to account for the various non-local dependencies that we find in Gs. Thus, Merge alone does not deliver what we need to fully explain the origins of our FL (i.e. it leaves out a large variety of agreement phenomena).[5] In this sense, either we need some  ideas about where AGREE comes from, or we need some work showing how to accomodate the phenomena that AGREE does via I-merge. Either way, the story that ties the evolutionary origins of our FL to the emergence of a single novel Merge operation is, at best, incomplete.

[1] Here from Edward St Aubyn in At Last: the final Patrick Melrose Novel:
 “Ok, so who created infinite regress.” That’s the right question.
[2] No less a figure than Wittgenstein had a field day with this observation.  “And so on” is not a concept that finite sequences of anything embody.
[3] I may be one of the last thinking that moving to AGREE systems was a bad idea if one’s interest is in DP. I argue this here. I don't think I’ve convinced many of the virtues either of disagreement in general or dis-AGREE-ment in particular. So it goes.
[4] It is tempting to see Chomsky’s latest discussions of labeling as an attempt to resolve this problem. Agreement on this view is what is required to get interpretable objects at the CI interface. It is not the product of AGREE but of the labeling algorithm. Chomsky does not say this. But this is where he might be heading. It is an attempt to reduce “morphology” to Bare Output Conditions. I personally am not convinced by his detailed arguments, but if this is the intent, I am very sympathetic to the project.
[5] I am currently co-teaching intro to contemporary minimalism with Omer Preminger. He has inundated me with arguments (good ones) that something like AGREE does excellent work in accounting for huge swaths of intricate data. Thus, at the very least, it seems that the current consensus among minimalist syntacticians is that Merge is not the only basic syntactic operation and so an account that ties all of our grammatical prowess to Merge is either insufficient or the current consensus is wrong. If I were a betting person, I would put my money on the first disjunct. But…

Thursday, April 23, 2015

Where the estimable Mark Johnson corrects Norbert (sort of)

For various reasons, Mark J could not post this as a comment on this post. As he knows much more than I do about these matters, I thought it a public service to lift these remarks from the comments section to make them more visible. I think that this is worth reading in conjunction with Charles' recent post (here). At any rate, this is interesting stuff and I don't disagree much (or have not found reasons to disagree much) with what Mark J says below. I will of, course,  allow myself some comments later. Thx Mark. 


This was originally written as a comment for the "Bayes and Gigerenzer" post, but a combination of a length restriction on comments and my university's not enabling blog posts from our accounts meant I had to email Norbert directly.

As Norbert has remarked, Bayesian approaches are often conflated with strong empiricist approaches, and I think this post does that too.  But even within a Bayesian approach, there are powerful reasons not to be a "tabula rasa" empiricist.  The "bias-variance dilemma" is a mathematical statement of something I've seen Norbert say in this blog: learning only works when the hypothesis space is constrained.  In mathematical terms, you can characterise a learning problem in terms of its bias -- the range of hypotheses being considered -- and the variance or uncertainty with which you can identify the correct hypothesis.  There's a mathematical theorem that says that in general as the bias goes down (i.e., the class of hypotheses increases) the variance increases.

Given this, I think a very reasonable approach is to formulate a model that includes as much relevant information from universal grammar as we can put into it, and perform inference that is as close to optimal as we can achieve from data that is as close as possible to what the child receives.  I think this ought to be every generative linguist's baseline model of acquisition!  Even with an incomplete model and incomplete data, we can obtain results of the kind "innate knowledge X plus data Y can yield knowledge Z".

But for some strange (I suspect largely historical) reason, this is not how Chomskyian linguists think of computational models of language acquisition.  Instead, they prefer ad hoc procedural models.  Everyone agrees there has to be some kind of algorithm which children use to learn language.  I know there are lots of pyschologists who are sure they have a good idea of the kinds of things kids can and can't do, but I suspect nobody really has the faintest idea of what algorithms are "cognitively plausible".  We have little idea of how neural circuitry computes, especially over the kinds of hierarchical representations we know are involved in language.  Algorithms which can be coded up as short computer programs (which is what most people have in mind when they say simple) might turn out to be neurally complex, while we know that the massive number of computational elements in the brain enable it to solve computationally very complex problems.  In vision -- a domain we can sort-of study because we can stick electrodes into animals' brains -- it turns out that the image processing algorithms implemented in brains are actually very sophisticated and close to Bayes-optimal, backed up with an incredible amount of processing power.  Why not start with the default assumption that the same is true of language?

It's true that in word segmentation, a simple ad hoc procedure -- a simple greedy learning algorithm that ignores all interword dependencies -- actually does a "reasonable job" of segmentation, and that improving either the algorithm's search procedure or making it track more complex dependencies actually decreases the overall word segmentation accuracy.  Here I think Ben Borschinger's comment has it right - sometimes a suboptimal algorithm can correct for the errors of a deficient model if the errors of each go in the opposite way.  We've known since at least Goldwater's work that an inaccurate "unigram" model that ignores inter-word interactions will prefer to find multi-word collocations and hence undersegment.  On the other hand, a naive greedy search procedure tends to over-segment, i.e., find word boundaries where there are none.  Because the unigram model under-segments, while a naive greedy algorithm over-segments, the combination actually does better than approaches where you just improve only the search procedure or only the model (by incorporating inter-word dependencies) since now you have "uncancelled errors".

Of course it's logically possible that children use ad hoc learning procedures that rely on this kind of "happy coincidence", but I think it's unlikely for several reasons.

First, these procedures are ad hoc -- there is no theory, no principled reason why they should work.  Their main claim to fame is that they are simple, but there are lots of other "simple" procedures that don't actually solve the problem at hand (here, learn words).  We know that they work (sort of) because we've tried them and checked that they do.  But a child has no way of knowing that this simple procedure works while this other one doesn't, so the procedure would need to be innately associated with the word learning task.  This raises Darwin's problem for the ad hoc algorithm (as well as other related problems: if the learning procedure is really innately specified, then we ought to see dissociation disorders in acquisition, where the child's knowledge of language is fine, but their word learning algorithm is damaged somehow).

Second, ad hoc procedures like these only partially solve the problem, and there's usually no clear way to extend them to solve the problem fully, so some other learning mechanism will be required anyway.  For example, the unigram+greedy approach can find around 3/4 of tokens and 1/2 of the types, and there's no obvious way to extend it so it learns all the tokens and all the types.  But children do eventually learn all the tokens and all the types, and we'll need another procedure for doing this.  Note that the Bayesian approach that relies on more complex models does have an account here, even though it currently involves "wishful thinking": as the models become more accurate by including more linguistic phenomena and the search procedures become more accurate, the word segmentation accuracy continues to improve.  We don't know how to build models that include even a fraction of the linguistic knowledge of a 3 year old, but the hope is that eventually these models would achieve perfect word segmentation, and indeed, be capable of learning all of a language.  In other words, there isn't a plausible path by which the ad hoc approach would generalise to learning all of a language, while there is plausible path for the Bayesian approach that relies on more and more accurate linguistic models.

Finally -- and I find it strange to be saying this to a linguist who is otherwise providing very cogent arguments for linguistic structure -- there really are linguistic structures and linguistic dependencies, and it seems weird to assume that children use a learning procedure that just plain ignores them.  Maybe there is a stage where children think language consists of isolated words (this is basically what a unigram model assumes), and the child only hypothesises larger linguistic structures after some "maturation" period.  But our work shows that you don't need to assume this; instead, a single model that does incorporate these dependencies combined with a more effective search procedure actually learns words from scratch more accurately than the ad hoc procedures.

Norbert sometimes seems very sure he knows what aspects of language have to be innate.  I'm much less sure myself of just what has to be innate and what can be learned, but I suspect a lot has to be innate (I think modern linguistic theory is as good a bet as any).  I think an exciting thing about Bayesian models is that they give us a tool for investigating the relationship between innate knowledge and learnability.  For example, if we can show that a model with innate knowledge X+X' can learn Z from data Y, but a model with only innate knowledge X fails to learn Z, then probably innate knowledge X' plays a role in learning Z.  I said probably because someone could claim that the child's data isn't just Y but also includes Y' and from model X and data Y+Y' it's possible to infer Z.   Or someone might show that a completely different set of innate knowledge X'' suffices to learn Z from Y.  So of course a Bayesian approach won't definitely answer all the questions about language acquisition, but it should provide another set of useful constraints on the process.