Grammar isn't merely part of language (2016)

http://www.dam.brown.edu/people/mumford/blog/2016/grammar.html

by mrkgnao • 9 years ago

124 68 9 years ago

dam.brown.edu

This article doesn't seem to make a strong case that a hierarchical tree is the best model for language (or motor control, or image recognition). Just because you can model something as a tree doesn't mean that's the most parsimonious or effective model. The relative lack of success thus far of symbolic parse tree based techniques in NLP compared to techniques grounded in other models should be a strong hint that trees are not the best map of the territory.

rntz• 9 years ago

I do not work on NLP, but my understanding was that, for purely syntactic work, standard parse-tree-based techniques had been quite successful in NLP; and that it is only for semantic work that symbolic representations begin to show weaknesses. Since we often care about the meanings of words, this is a pretty strong limitation; still, it suggests that standard grammar-and-parse-tree approaches do capture something significant about how human languages work.

Is this inaccurate?

smallnamespace3• 9 years ago

I don't work in NLP either, but from my understanding, the boundary between syntax and semantics is never as clean as one might imagine, and each language draws the boundary differently, so in general the utility of just looking at syntax can greatly vary.

Another issue is that many real life sentences have more than one possible parse, and we use context and semantics to disambiguate, e.g. how do you parse 'fruit flies like a banana'.

YeGoblynQueenne13• 9 years ago

The problem with what you're saying (that grammars are not necessarily the best representations of complex hierarchical structures, with which I agree) is that anything that can represent a complex hierarchical structure as well as a grammar must necessarily be equivalent in computational power to that grammar- and unfortunately, we know of no computational process that cannot be expressed as a grammar at least in principle.

"Unfortunately" because this means that any representation of such a process you may want to pick over a gramar, that is not a grammar, will either have to be reducible to a grammar, or fail to capture the expressive power of the modelled process.

So, sure, a "tree" may not be the best way to model natural language. But we don't have the theoretical tools to figure out what can be a better representation than that.

mannykannot• 9 years ago

Is this not almost tautologically true? Will not any collection of related information have the potential to be represented by some sort of hierarchical structure?

stonemetal• 9 years ago

It can be represented as a hierarchical structure as long as the relationships are simple. In any of his examples if you added a number of extra relationships to complicate what is there, the hierarchical structure becomes a generic graph.

raverbashing• 9 years ago

While we like to think of Grammar as a perfect logical structure, in practice, "the brain" (or some higher level structure) processes things in a highly probabilistic fashion. This is how we are able to understand imperfect information (and the mention of the animal "grammar" is on point - a pet can learn and find their way around a house easily)

It's probably something more complex than a LTSM structure, but the point stands.

moomin• 9 years ago

Hasn't Chomsky's idea been fairly comprehensively disproven? I seem to recall a survey of language features that concluded there were no clusters of grammatical constructions of the form his theory predicts.

txru• 9 years ago

No, not at all. I think people might be surprised at just how basic the structures that he uses his evidence are.

The Piraha study was a study with poor research practices (subsequent researches have found conflicting findings from Everett's), and always ignored one massive problem, as told by a joke:

"Professor Chomsky! There's a language in Brazil that doesn't seem to have recursion or large numerals!"

Chomsky doesn't turn around.

"Can they learn Portuguese?"

"... Yes."

That's a lot of what his theories are-- that for a set of features in human language including types of structure like recursion, every homos sapiens has the ability to use them in language.

bambax4• 9 years ago

Not a very good joke. Everett doesn't argue that the Pirahas are a different species; they're human so of course they can do what humans do!

What he's saying (and I think demonstrated pretty well) is that not ALL languages are based on the same fundamental structures: to show this, one counterexample is enough.

joe_the_user• 9 years ago

Yeah no,

The problem is that language is not simply a hierarchical representation of reality. Human language involves a number of uniqe specific qualities - enough that I can't list all of them off the top of my head but just for example, symbolic representation.

Sure, an animal represents the world in its brain. But a lot of animals don't get that a symbol in the world can represent another symbol in the world. I know a common example is pointing - only the most representation-adept animals "get" that a person can use their finger to indicate something else.

And this, symbolic representation, is just one aspect of human language broadly.

woodandsteel2• 9 years ago

Why do linguists like Chomsky assume that grammar is a completely radical break from animal brain functioning? I think it is mainly because of a deeply rooted assumption in Western religion and much (but not all) of Western philosophy that human beings have free, moral, non-material souls that animals lack.

robertk• 9 years ago

It clearly can't be. The genetic SNP delta between humans and other primates is too small to record sufficiently complex cortical changes for something like all of neural symbolic representations.

nitrogen• 9 years ago

Even simple cellular automata can produce drastically different behavior given slight tweaks to simple rules, so I don't think that a small genetic delta rules out a large structural change.

slurry• 9 years ago

Name an animal that uses grammar, or something like it.

empath75• 9 years ago

This blog post gives examples.

sunstealer2• 9 years ago

Chomsky's reasoning is that the arbitrary complexity (recursive structure) of human language implies some sort of low level computational engine to do the relevant computations. I don't have an opinion either way. We can see that LSTMs can emulate this kind of logic, but they also make mistakes. Also, I'm not sure that human reasoning is as logical as it might seem. E.g. I read somewhere (lost the reference) that the earliest languages may have lacked the ability to arbitrarily nest clauses. So maybe humans only emulate logical thinking.

visarga• 9 years ago

arbitrary complexity (recursive structure) of human language

Interesting, because I was reading this paper yesterday, where they argue that recursivity is the key to generalization power.

"Making Neural Programming Architectures Generalize via Recursion" - https://openreview.net/forum?id=BkbY4psgg&noteId=BkbY4psgg

Al-Khwarizmi2• 9 years ago

We can see that LSTMs can emulate this kind of logic, but they also make mistakes.

Humans can also make mistakes when processing language. We're still better than LSTM's, but I'm not sure we can claim a qualitative difference.

Furthermore, even though we can process sentences with very deep embedding like "The rat the cat the dog bit chased escaped", my intuition is that we are not using our normal language processing systems for that. When I read that sentence, I just fail to process it and then I invoke my logic systems to try to determine the structure and decode it, in a way that feels totally different from processing a normal sentence (I'm not understanding it in real time, in a natural way, but rather solving a small puzzle). So I personally don't find the Chomskyan arguments based on that kind of corner cases very convincing.

agumonkey3• 9 years ago

What annoys me about linguistics is that higher level languages are quite poor for basic human communication. Non verbal communication is maybe the most important channel.

Typhon3• 9 years ago

Actually human languages are much better suited for all kinds of communication : To the best of our knowledge, most animal systems of communication don't allow reference to either the speaker (i.e "I"), the hearer (i.e "you"), or a distant third party (i.e "it"). Even elaborate systems like the bee dance are highly specialized in the type of reference that is possible. In human speech it is possible to talk about things or people that are not here, things that don't exist but may, things that have existed, things that will exist, things that won't exist. Hypotheses, conjectures, and so on, are not part of any other species way of communicating, as far as I know.

Furthermore, all known human languages allow you to attribute utterances to another, earlier speaker, what is known as reported speech (often, though not always as Pirahâ demonstrate, through an embedded clause "He said that X").

Human non-verbal communication (which is EMPHATICALLY NOT to be confused with sign languages), on the other hand, is typically not very expressive. Apart from the admittedly very important, but nevertheless quite specialized domain of human emotion (conveyed by facial expression for instance), it is more or less inadequate for much of what we want to communicate.

Typhon3• 9 years ago

Not every linguist agree with Chomsky.

eternalban3• 9 years ago

I share your suspicions.

non-material souls

I think this notion of "materiality" is an ancient chain letter that we can discard, or at least leave aside as we consider alternatives that can still support the notional 'self' & individual 'soul' without requiring a dualistic reality.

What the ancients understood as "matter" we can consider as 'localized coherence' (membranes, knots, vortices, etc.), and the "spirit" as fields, in a non-dualistic universe.

So speaking of Western religions, reconsider the meaning of the thought that "the kingdom is in you", and in the spirit of a holistic human metaphysics, note that whether in the "East" or in the "West", the key to transcendence appears to be shedding the illusion of the centrality of the localized "body" as the totality of the self.

d_burfoot• 9 years ago

I'm a huge fan of Mumford, but I think he's stretching a bit when he says that natural images have grammar like language has grammar. The production processes for the two phenomena are enormously different. A natural image is formed when a collection of objects is illuminated by incoming light, and the resulting image is projected onto the retina. The human brain is not involved at all in this process (leaving aside nitpicking about how humans may have shaped the environment). In contrast a natural language sentence is produced when an idea occurs inside the brain, and then various linguistic production processes transform the idea into a serial form, as text, speech, sign language, etc. The latter process involves constraints, capabilities, and eccentricities of the human brain at every stage.

Maybe you could argue that human brains perceive images using grammar-like structures.

yellowapple• 9 years ago

Human brains also perceive speech or writing or what have you using grammar-like structures. In the sense brought up by the article, a grammar is just a structured logical representation of some physical phenomenon, like sound or imagery.

In other words, grammar is how we perceive things, and it makes sense that it can be generalized instead of only being applicable to a specific sensory input.

The specific point about parsing visual input is much more obvious when looking at things like graphical charts, user interfaces, etc.; there's clearly a grammar of some form involved in, say, determining which button to press on my phone's on-screen keyboard to create the letter 'b', and further involved when determining what a "button" is or what a "keyboard" is or what a "letter" is. Hell, that seems like the same process that lets me figure out what a "screen" is and what a "phone" is. Eventually we go from "type the letter 'b'" to "move your right thumbtip to this position" (and even that can be broken down further).

mistersys• 9 years ago

You're treating input/output as the same.

We receive speech coming from a source and parse it using a grammar. One could imagine it being a similar process for the perceiving images captured by the retina.

For output, when a human paints an image they are panting from an imaged visualized inside of their mental canvas, just like we realize thoughts produced within our minds as speech.

smallnamespace3• 9 years ago

when a human paints an image they are panting from an imaged visualized inside of their mental canvas

Yes, but images produced by humans are a tiny fraction of images processed by the eye. But every written or spoken sentence was ultimately created by a human brain.

That's why it seems like a big stretch to claim there is a 'universal grammar' involved in visual processing, if you believe that grammar is primarily a way for brains to encode information for communication purposes...

jacobolus• 9 years ago

I highly recommend you read Arnheim’s 1969 book Visual Thinking, https://amzn.com/0520013786/

YeGoblynQueenne13• 9 years ago

> I'm a huge fan of Mumford, but I think he's stretching a bit when he says that natural images have grammar like language has grammar.

That seems to be what he's saying- that the process of vision in living things is actually a grammar.

He doesn't have to be entirely correct for his intuition to be that useful though. We don't have to assume an existing grammar to fit a bunch of data to a grammar (like we do in grammar induction, where there's no such assumption when, say, someone models DNA sequences as a grammar etc).

After all, a grammar is just a representation. The question is how good that representation is- in theory as well as in practice. In theory, it's a good representation if it helps us answer questions about the process we're trying to model. In practice it's good if it allows us to reproduce the process, especially automatically, with computers, and predict the behaviour of agents that employ this process etc etc.

eternalban3• 9 years ago

Recursion was recently shown to enable generalization in neural programming architectures [1] but from a critical inquiry p.o.v. we note that recursion requires a mechanism for maintaining context and look for the existence (or lack of) such mechanism in animal brains.

[1] https://news.ycombinator.com/item?id=13551298

clord• 9 years ago

A natural image is formed when a collection of objects is illuminated by incoming light, and the resulting image is projected onto the retina.

But that's sort of the point no? The incoming information is not an unstructured white noise of photons striking our retina. There is a sort of structure to the information that can be modelled. One such model of this structured information is as a "grammar tree" (really just a tree, we're coders here.) The example in the article is that the arm occludes the teepee, which occludes the background trees. Any visual system needs to break this hierarchy down.

woodandsteel2• 9 years ago

A natural image is formed when a collection of objects is illuminated by incoming light, and the resulting image is projected onto the retina. The human brain is not involved at all in this process

No, the retina is a complex processor, and so is the optic nerve. Brain scientists nowadays say the retina is an extension of the brain.

bambax4• 9 years ago

the feud between Noam Chomsky and Daniel Everett

"Don't sleep, there are snakes" which recounts the experience of Everett among the Pirahas people is a fantastic book.

ktRolster2• 9 years ago

Pretty sure Chomsky won that 'fight' hands down.

bambax4• 9 years ago

Pretty sure he didn't. Also, Everett book isn't only about the controversy with Chomsky, it's about his own experience in the jungle and his coming of age. It's a great book even if you absolutely love Chomsky.

hyperpallium• 9 years ago

By recursion Chomsky means of the same component, which is potentially infinite. Hierachical and recursion is like the difference between regular expressions and context free grammars.

Of course, in practice, most people struggle with just a few levels of recursion. So I'm not sure I agree with Chomsky - but that's his theory.

BTW the grammar-related Piraha ideology of immediate experience is appealing, as theories and abstractions can obscure what's actually happening, and reminds me of DNA's ruler of the universe.

paufernandez• 9 years ago

Thanks a lot for posting this! It made me very happy. I hold the same opinion as him and for many years I've searched the literature looking for people with similar views!

YeGoblynQueenne13• 9 years ago

The problem with representing processes with grammars is that you can represent pretty much any process with a grammar, right up to the level of Turing machines and the Lambda calculus etc. The question is, always: how useful is it to do this in practice?

With language, we assume that there is some sort of underlying structure that can best be modelled as a grammar (well, some of us do). In simple terms, we assume that natural language already has a grammar so that we can hope to eventually reconstruct it somehow- either by hand, or by automated inference etc.

Unfortunately, in practice every effort to do that sort of thing undertaken since Chomsky's Syntactic Structure was published in 1957, has been met with failure and even theoretically the outlook is bleak (see Gold's result and a ton of bibliography on inductive inference before and after it).

In short, the problem is not that it's hard to convince ourselves that there are processes in nature that are best modelled using hierachical representations, like grammars. The problem is that even armed with those representations we've so far proven incapable, with all our science and technology, to accurately use such representations to fully model those processes.

Basically, our current situation regarding the modelling of complex hierarchical processes is like being given the key to heaven, but no map to the damned gates.

eternalban3• 9 years ago

Basically, our current situation regarding the modelling of complex hierarchical processes is like being given the key to heaven, but no map to the damned gates.

More accurately, you have been given a key that is claimed to be the key to heaven, but no one has yet found a gate that the given key opens, yet you insist on asserting that it is the key to "the damned gates".

YeGoblynQueenne13• 9 years ago

Ouch.

Look, no. Grammars look like "the key to the gates" because of the equivalence between grammars-language-automata etc. You can write up grammars for fully Turing-complete languages, or indeed context-free languages that allow you to declare Turing-complete automata.

You can create grammars that display infinite recursion with minimal effort:

  A --> ε
  A --> Aa
  a --> <whatever you please>

So it's not like someone (cough, Chomsky) woke up one nice day and thought "blimey, I'll tell the world that grammars are a powerful tool for modelling hierarchical processes". It's that they are.

We have used grammars in practice to model complex hierarchical processes- except it's only those processes that we already know how to model, because we came up with them ourselves, like the aforementioned Turing-complete ones.

The problems begin when we try to fit a process we hardly understand to a grammar. That is not a limitation of the tool itself. It's a limitation of our ability to use it.

YeGoblynQueenne13• 9 years ago

> the feud between Noam Chomsky and Daniel Everett

That "feud" is as much a feud as the "controversy" over climate change. Daniel Everett is very obviously a complete and utter crank, who has made absolutely ridiculous claims about Piraha the language and Piraha the people (e.g. that they can't learn basic arithmetic skills, like 1 + 1), to which he claims to be the sole authority.

To give an account of his "feud" with Chomsky, Chomsky claims that recursion (in the sense of embedding sentences into sentences) is the defining characteristic of human language that sets it apart from other animal languages. Everett claims that the Piraha language doesn't display recursion and therefore Chomsky's claim is wrong.

Stop for a moment and consider this. It's like claiming that, only Europeans have pale skin, but if we find one European population with brown skin, then it's not true that only Europeans have pale skin. What Everett is claiming is exactly that wrong.

Everett has only risen to prominence because he's a big old troll, and there are lots of people in cognitive science and linguistics who are very frustrated with the inability of both fields to make any progress on natural language, despite everyone's best efforts- and who have somehow singled out Chomsky as the culprit, because he had the audacity to hold on to a theory that hasn't really been comprehensively disputed yet.

Far from a feud, this is just a case of shoot-the-messenger. Shoot him with sour grapes that is.

canjobear2• 9 years ago

The other side isn't exactly a bunch of careful scientists either [0].

Other people have visited the Pirahã and done experiments with them. It is true that they do not understand counting, for example [1][2], though I'm sure they could learn it if they had to.

Other linguists have also independently claimed that other languages lack recursive embedding, for example Riau Indonesian, so the claim about Pirahã is not isolated or crazy a priori.

The Chomskyan side of the argument has also been characterized by moving the goalposts and formal incoherence. See for example [3].

Basically, amidst a huge storm of verbiage and anger and inflated claims from both sides, all of which is about personalities and definitions and not at all about science, there are a few mildly interesting findings about Pirahã and language.

[0] https://en.wikipedia.org/wiki/Marc_Hauser [1] http://science.sciencemag.org/content/306/5695/496.full [2] http://lchc.ucsd.edu/mca/Mail/xmcamail.2014-12.dir/pdf2Yb7JA... [3] http://www.sciencedirect.com/science/article/pii/S0388000114...

YeGoblynQueenne13• 9 years ago

> Other people have visited the Pirahã and done experiments with them. It is true that they do not understand counting, for example [1][2],

Those are not "other people". The first reference is a paper where Everett is one of the authors and the other is a paper from a collaborator of Everett: the author even studied the Piraha and their arithmetic skills while living with the Piraha along with the Everetts. Show me a paper on the Piraha language that has nothing to do with Everett and that duplicates Everett's findings.

> The Chomskyan side of the argument has also been characterized by moving the goalposts and formal incoherence. See for example [3].

I don't think that reference means what you think it means. Actually I think it means exactly the opposite:

What I can say, what I have said, with most certainty is what recursion is supposed to stand for within the theory Chomsky has constructed—with remarkable consistency, I should add.

Which is what I've also heard before from critics of Chomsky in general (frex, I believe Alex Clark has said similar things, but I might be wrong). Chomsky himself has been remarkably consistent, to the point he sounds like a broken record, regarding what recursion in language means.

Whether "his side" (other linguists of his school) has been more or less consistent, I don't know but in any case I'm more interested in the fact that Daniel Everett seems to be a complete and utter sleazebag who is unashamedly taking advantage of a peoples who can't really speak for themselves (because he's standing in the way!) to further some obscure personal agenda.

> Other linguists have also independently claimed that other languages lack recursive embedding, for example Riau Indonesian, so the claim about Pirahã is not isolated or crazy a priori.

I hadn't heard of Riau Indonesian, so I looked it up in wikipedia and I could not find anything about it lacking recursive structure. Instead it "is considered by linguists to have one of the least complex grammars among the languages of the world,[citation needed] apart from creoles, possessing neither noun declensions, temporal distinctions, subject/object distinctions, nor singular/plural distinction."

YeGoblynQueenne13• 9 years ago

I'll quote from your last reference:

As I have tried to show, those issues, whilst important and interesting in themselves, have little to do with recursion as this concept has been understood within mathematical logic, the connotation Chomsky has always employed, it seems. As such, there is a bit of a disconnect between what Chomsky is talking about when talking about recursion and what other linguists are talking about when talking about recursion. As a result, rather incompatible claims are being put forward in the literature; indeed, to defend that what is universal in linguistics is the recursively-defined operations of a computational system (clearly Chomsky's take) is very different from the claim that all languages must exhibit self-embedded sentences if recursion is to be accepted as such a central concept (Everett is probably the most prominent example of this line of thought).

Like I say- that paper basically claims that Chomsky's version of recursion in language has changed very little if at all, and that Everett is grossly misrepresenting it (or its consistency).

I would add: willfully so.

mcguire3• 9 years ago

"To give an account of his "feud" with Chomsky, Chomsky claims that recursion (in the sense of embedding sentences into sentences) is the defining characteristic of human language that sets it apart from other animal languages. Everett claims that the Piraha language doesn't display recursion and therefore Chomsky's claim is wrong."

I realize that your summary of Chomsky's theory is short, pithy, and misrepresents Chomsky, but your argument here is incorrect.

You say,

* X is the defining characteristic of Y.

* Z (a purported element of Y) does not have X.

The only possible conclusion is that Z is not an element of Y: the Piraha do not speak a human language.

YeGoblynQueenne13• 9 years ago

Sure, that's Everett's (wild, unsubstantiable, absurd) claim: that the Piraha, a human people, speak a nonhuman language.

And that affects the claim that recursion is the defining characterstic of human language -how, again?

Grammar isn't merely part of language (2016)

Related Stories