Whilst browsing through the language acquisition page on Wikipedia, I noticed a small segment labelled "Statistical Learning":

Statistical learning
Some language acquisition researchers, such as Elissa Newport, Richard Aslin, and Jenny Saffran, believe that language acquisition is based primarily on general learning mechanisms, namely statistical learning. The development of connectionist models that are able to successfully learn words and syntactical conventions[19] supports the predictions of statistical learning theories of language acquisition, as do empirical studies of children's learning of words and syntax.[20]

After reading more on it (e.g. here), I struggled to find where it differs from the behaviorist theories on language. Anybody able to help me figure out what I'm missing?

I've read the pdf you've linked and it seems it might be coming from a different angle (?) but to essentially compatible conclusions. Interesting read, and I liked the handling of innate language "knowledge" vs. learning debate wrt language universals. Not that this is a revolutionary result, but the point needs to be repeated and supported and hammered home ...

Someone sent me this:

---
At TEDxRainier, Patricia Kuhl shares astonishing findings about how babies learn one language over another -- by listening to the humans around them and "taking statistics" on the sounds they need to know. Clever lab experiments (and brain scans) show how 6-month-old babies use sophisticated reasoning to understand their world.

http://www.ted.com/talks/patricia_kuhl_ ... abies.html
---

few days ago. Convincing. Reasonable. Would it be surprising if language acquisition would be multi-dimensional?

---
edit: BTW, the ability, of babies, is lost forever so to say, cannot be learned/regained ever?

katja z wrote:I've read the pdf you've linked and it seems it might be coming from a different angle (?) but to essentially compatible conclusions.

Are you able to expand on this? I'm just having trouble figuring out what the different angle is, because behavioral theories rely on a statistical distribution of words and phonemes in order for certain meanings (and grammars etc) to be reinforced. This kind of connectionist network that results is exactly what the behavioral models propose.

katja z wrote:Interesting read, and I liked the handling of innate language "knowledge" vs. learning debate wrt language universals. Not that this is a revolutionary result, but the point needs to be repeated and supported and hammered home ...

I think one of the "differences" is that the statistical theories seem to explicitly state the dual role of nature and nurture, whereas the behavioral models tend to take it for granted, resulting in the misconception that it's a blank slate approach to language acquisition.

tuco wrote:Someone sent me this:

---
At TEDxRainier, Patricia Kuhl shares astonishing findings about how babies learn one language over another -- by listening to the humans around them and "taking statistics" on the sounds they need to know. Clever lab experiments (and brain scans) show how 6-month-old babies use sophisticated reasoning to understand their world.

http://www.ted.com/talks/patricia_kuhl_ ... abies.html
---

few days ago.

Thanks for that, tuco

tuco wrote:Convincing. Reasonable. Would it be surprising if language acquisition would be multi-dimensional?

I don't think it would be surprising at all, and it must be inevitable, in my opinion.

tuco wrote:edit: BTW, the ability, of babies, is lost forever so to say, cannot be learned/regained ever?

Are you referring to the ability to pick up the phonemes from any language? As far as I know, the flexibility and rapid ability to pick it up is lost forever, but of course later training can help people pick up a second language. I think this is because once we have a basic framework for language (i.e. the neural network structure), further language learning builds from that starting point. This means that we essentially have to forget some things we knew, and what we've learnt can interfere with learning new things (e.g. two phonemes which sound similar but are pronounced differently).

Mr.Samsa wrote:

katja z wrote:I've read the pdf you've linked and it seems it might be coming from a different angle (?) but to essentially compatible conclusions.

Are you able to expand on this? I'm just having trouble figuring out what the different angle is, because behavioral theories rely on a statistical distribution of words and phonemes in order for certain meanings (and grammars etc) to be reinforced. This kind of connectionist network that results is exactly what the behavioral models propose.

Hmm on second thoughts, "angle" was probably not the best of words. A different focus might be better. What I meant was - the article you linked does reference general learning mechanisms and to that extent is perfectly compatible with how a behaviourist would approach the issue. But it seems to be more focussed on the language than on the learner; by that I mean, more focussed on the detectable statistical properties of what is to be learned, than on the mechanisms of reinforcement. Hope this helps.

(Also, vocabulary analysis seems to suggest the author is not working within a behaviourist framework )

tuco wrote:edit: BTW, the ability, of babies, is lost forever so to say, cannot be learned/regained ever?

Are you referring to the ability to pick up the phonemes from any language? As far as I know, the flexibility and rapid ability to pick it up is lost forever, but of course later training can help people pick up a second language. I think this is because once we have a basic framework for language (i.e. the neural network structure), further language learning builds from that starting point. This means that we essentially have to forget some things we knew, and what we've learnt can interfere with learning new things (e.g. two phonemes which sound similar but are pronounced differently).

Good points. I'd like to add that phonemes are just one layer of language. To a certain extent, children do seem to lose the ability to "hear" phonetic distinctions not relevant to their linguistic environment as a matter of synaptic pruning.
[speculation] On the other hand, this doesn't necessarily affect other areas of language learning, for instance recognition and use of complex syntactic patterns. On the contrary, previous experience with these might well speed up the process (we shouldn't forget that interference with (a) previously acquired language(s) is not always detrimental, it can be helpful too, since the knowledge of certain features can be simply transferred and adapted across a number of languages). [/speculation]

katja z wrote:

Mr.Samsa wrote:

katja z wrote:I've read the pdf you've linked and it seems it might be coming from a different angle (?) but to essentially compatible conclusions.

Are you able to expand on this? I'm just having trouble figuring out what the different angle is, because behavioral theories rely on a statistical distribution of words and phonemes in order for certain meanings (and grammars etc) to be reinforced. This kind of connectionist network that results is exactly what the behavioral models propose.

Hmm on second thoughts, "angle" was probably not the best of words. A different focus might be better. What I meant was - the article you linked does reference general learning mechanisms and to that extent is perfectly compatible with how a behaviourist would approach the issue. But it seems to be more focussed on the language than on the learner; by that I mean, more focussed on the detectable statistical properties of what is to be learned, than on the mechanisms of reinforcement. Hope this helps.

(Also, vocabulary analysis seems to suggest the author is not working within a behaviourist framework )

Ah, that makes sense! Thank you, I think I understand now

katja z wrote:Good points. I'd like to add that phonemes are just one layer of language. To a certain extent, children do seem to lose the ability to "hear" phonetic distinctions not relevant to their linguistic environment as a matter of synaptic pruning.
[speculation] On the other hand, this doesn't necessarily affect other areas of language learning, for instance recognition and use of complex syntactic patterns. On the contrary, previous experience with these might well speed up the process (we shouldn't forget that interference with (a) previously acquired language(s) is not always detrimental, it can be helpful too, since the knowledge of certain features can be simply transferred and adapted across a number of languages). [/speculation]

Yes that is a good point - learning the basics of grammar, for example, can help to understand how the concept of grammar functions across all languages (even if the form it takes differs from one's original language).

Mr.Samsa wrote:After reading more on it (e.g. here), I struggled to find where it differs from the behaviorist theories on language. Anybody able to help me figure out what I'm missing?

Here you have an interesting article that compares the behavioral and cognitive views on speech perception and production, including statistical learning:
http://www.thefreelibrary.com/Behaviora ... 0229835130

seeker wrote:

Mr.Samsa wrote:After reading more on it (e.g. here), I struggled to find where it differs from the behaviorist theories on language. Anybody able to help me figure out what I'm missing?

Here you have an interesting article that compares the behavioral and cognitive views on speech perception and production, including statistical learning:
http://www.thefreelibrary.com/Behaviora ... 0229835130

Thanks Seeker, that is an excellent article. Out of interest, where do you stand on the issue of statistical learning? Is it something that can be explained by operant conditioning principles, or is it something unique to the structure of language itself?

Mr.Samsa wrote:Thanks Seeker, that is an excellent article. Out of interest, where do you stand on the issue of statistical learning? Is it something that can be explained by operant conditioning principles, or is it something unique to the structure of language itself?

I think that statistical learning is more closely related with pavlovian conditioning principles (and not with operant conditioning principles), because statistical learning is the effect of pavlovian contingencies between stimuli (and not of operant contingencies between responses and their consequences).
I don't think it's unique to the structure of language. As far as I know, there's evidence of statistical learning in babies that have not yet learned a language.

Mr.Samsa wrote:Whilst browsing through the language acquisition page on Wikipedia, I noticed a small segment labelled "Statistical Learning":

Statistical learning
Some language acquisition researchers, such as Elissa Newport, Richard Aslin, and Jenny Saffran, believe that language acquisition is based primarily on general learning mechanisms, namely statistical learning. The development of connectionist models that are able to successfully learn words and syntactical conventions[19] supports the predictions of statistical learning theories of language acquisition, as do empirical studies of children's learning of words and syntax.[20]

After reading more on it (e.g. here), I struggled to find where it differs from the behaviorist theories on language. Anybody able to help me figure out what I'm missing?

Thanks for the links ! I should visit the Linguistics section more often. As it happens I am quite partial to the statistical learning hypothesis (even if perhaps not as an all-encompassing theory of language). This is because I have seen it (read it, I mean) doing wonders in explaining how music works.

As for fitting it within an operant conditioning framework, it is certainly possible with some extra assumptions. Suppose that when listening to speech we always engage into predicting the speaker's next utterance (there are some hints that we do that, like the ease with which we complete unfinished sentences, and the strong immediate surprise that comes with an unexpected word). Since learning language is adaptive, it makes sense that correct predictions would be rewarded. Moreover, our predicting competence would enable us to speak instead of privately predicting. I guess that with the right values fit in here and there, the matching law would imply statistical learning.

Now the above seems to have more explanatory value for music, which is none too adaptive, and where the #1 puzzle is to explain why we enjoy music so much in the first place. In the case of language, one could argue that the private predicting behavior thus posited is unparsimonious, and that "naked" statistical learning is a more parsimonious explanation. Or in a cognitive wording, that statistical learning works at a lower level than operant conditioning, so is more fundamental. I am not well-versed enough into the philosophy of mind to decide on that, but it is always nice to see when different theories say more or less the same thing.

Mr.Samsa wrote:Whilst browsing through the language acquisition page on Wikipedia, I noticed a small segment labelled "Statistical Learning":

Statistical learning
Some language acquisition researchers, such as Elissa Newport, Richard Aslin, and Jenny Saffran, believe that language acquisition is based primarily on general learning mechanisms, namely statistical learning. The development of connectionist models that are able to successfully learn words and syntactical conventions[19] supports the predictions of statistical learning theories of language acquisition, as do empirical studies of children's learning of words and syntax.[20]

After reading more on it (e.g. here), I struggled to find where it differs from the behaviorist theories on language. Anybody able to help me figure out what I'm missing?

Skinner's model and the statisitical learning model are both learning models. They are both empiricist, if that's what you're asking.

Mick wrote:Skinner's model and the statisitical learning model are both learning models. They are both empiricist, if that's what you're asking.

Nah, I wasn't just commenting on the similarities, but rather I was suggesting that they appeared to be the same thing. Katja, Palindnilap, and Seeker have helped me figure out where they diverge though.

Well that's a term i hadn't heard before. I had a quick look at the article and then did a google on statistical language learning. Not really helpful as nobody seems to want to define their terms succinctly. This is typical of many wooly concepts in the Psycholinguistic literature of early child language acquisition I fear.

What I did seem to glean from this quick search is that this area MIGHT be emphasising learning by exposure alone. Whereas a Behavioural approach would emphasise the importance of reinforcement in shaping language learning.

The video that tuco posted of Patricia Kuhl seems to be saying that babies will learn foreign sounds by some mystical ability to "take statistics" of the sounds by exposure to them. She then goes on to highlight how important it is for there to be a human present to speak these words to the baby. So they don't learn the new sounds from exposure alone then ....duh!!!!

When a human is present in these sorts of studies, as it is in Kuhl's video it is impossible to control for the effects of reinforcement by the researcher/reader of words. So any inferences that they are learning via some complicated mental reasoning strategy, or indeed by some innate and mysterious language acquisition device, can be taken with a pince of salt