Popular in Psychology

Most Cited

The adapted mind

Dalhousie University Jerome Barkow

Psychology

Consumers’ pursuit of material and experiential purchases

Cornell University Tom Gilovich

Experimental Psychology

Time to investigate human visual shape perception

University of Montreal Martin Arguin

Most Cited

James S. Jackson and the program for research on Black Americans

Harvard University David R. Williams

Psychology

Violence and Mental Illness: Rethinking Risk Factors and Enhancing Public Safety

Duke University Eric Elbogen

Cognitive Psychology

Cognitive remediation – where are we now and what should we do next?

King's College London Til Wykes

Cognitive Psychology

Harnessing Plasticity in the Atypical Developing Brain using Neurostimulation

University of Surrey Roi Cohen Kadosh

The Acquisition of Noun and Verb Categories by Bootstrapping From a Few Known Words

View more insights in Social Sciences - Psychology - Language and Memory

View more figures from University of Liverpool

View the page of Perrine Brusini

Read the Study

Next in Language and Memory

Metaphorical language processing and amygdala activation in non-native and native speakers

Francesca Citron | Lancaster University

Who’s a good boy?! Dogs prefer naturalistic dog-directed speech

Alex Benjamin | University of York

While many studies have shown that toddlers are able to detect syntactic regularities in speech, the learning mechanism allowing them to do this is still largely unclear. Perrine Brusini discusses computational modeling to assess the plausibility of a context-based learning mechanism for the acquisition of nouns and verbs.

Image courtesy of interviewee. May 1, 2024

Log-in or Sign-up to Faculti
Currently viewing this subject insight as a guest. You have insight(s) remaining for this month. Login to view 8000+ figures on the platform.

Generating Workspace...

Export as PDF

This video video is a transcript from a presentation discussing the complexity of human language, particularly focusing on syntax, which is the ability to combine words following certain rules, often unconsciously. The speaker emphasizes that this ability is unique to humans and is referred to as generative capacity, allowing individuals to understand and produce sentences they have never encountered before.

The discussion touches on the development of language in infants, concerning when they start using these complex rules. The speaker references Noam Chomsky’s theory that linguistic abilities are biologically innate, which has sparked much debate in the field. Extensive research has been conducted to determine at what age infants can manipulate syntax, with challenges in differentiating between grammatical sentences and those that simply have low probabilities of word co-occurrence.

The research presented involves modeling experiments to assess how infants might acquire understanding of nouns and verbs through the input they receive from caregivers and their environment. The modeling approach uses naturalistic data from caregiver-infant interactions, analyzing vocabulary and the relationship between context and word categories.

Results from these models show that they can effectively categorize nouns and verbs, with relatively high precision rates, demonstrating a potential method through which infants might learn syntax using limited vocabulary and semantic cues. The audience is informed about future research directions, including verifying if infants can indeed recognize semantic categories in a way similar to adults and how they apply this understanding when encountering new words. Overall, the presentation explores significant questions around language acquisition and the cognitive processes involved in learning syntax.

Language:

Export as PDF

Export as PDF
I'd be happy to help! However, it seems that you didn't provide the specific video you want me to analyze. Please share the video you're referring to, and I'll extract the five key points for you.

Export as PDF
It seems you have not a specific video to analyze for key quotes. Please provide the video or specify the work you are referring to, and I would be happy to help you find key quotes from it.

X-TIMESTAMP-MAP:LOCAL=00:00:00.000,MPEGTS=900000

00:00.075 --> 00:03.685
When we use language, we, we don't really realize it

00:03.685 --> 00:06.485
because it's very automatic, for us adults,

00:06.585 --> 00:08.325
but we use many different units.

00:08.865 --> 00:11.245
And one of, the important one,

00:11.265 --> 00:15.205
and that has been argued to be, the facts

00:15.355 --> 00:18.445
that we are specific language is specific to human

00:19.145 --> 00:20.405
is, syntax.

00:20.545 --> 00:23.245
So syntax is our capacity to combine war together.

00:23.745 --> 00:26.805
And what is interesting is that we combine,

00:26.805 --> 00:28.605
words together according to rules

00:28.875 --> 00:30.445
that we are not totally aware.

00:30.745 --> 00:33.365
But there is sentence that you will never produce

00:33.505 --> 00:35.205
as a speaker of your own language.

00:35.755 --> 00:40.565
Some things like, flower on the,

00:40.705 --> 00:42.685
yellow, that doesn't make any sense.

00:42.705 --> 00:45.405
You cannot do this and you will never produce it.

00:45.875 --> 00:47.925
What is interesting also is that

00:47.925 --> 00:50.325
because you know, some of the rule of your language,

00:50.745 --> 00:53.925
you are able to integrate sentences

00:53.925 --> 00:55.365
that you have never heard before.

00:55.785 --> 00:59.565
So if I say something like, the cat is mm-hmm.

00:59.775 --> 01:00.845
Under the mm-Hmm.

01:01.355 --> 01:05.605
Even if you don't know at all what those non-word that is,

01:05.775 --> 01:09.325
-mm are you start to understand what is happening.

01:09.385 --> 01:11.805
You have a fairly good idea of the action

01:11.805 --> 01:14.845
that the cat is doing and where is doing it.

01:15.545 --> 01:19.005
So, this capacity seems to be very specific to human.

01:19.275 --> 01:20.485
It's called generative.

01:20.905 --> 01:21.925
So we can generate

01:22.265 --> 01:24.965
and decode sentences that we have never heard.

01:25.345 --> 01:29.005
And so one of the question that is crucial for my work is

01:29.005 --> 01:33.165
to know from when infants can do that, when do we start

01:33.265 --> 01:37.205
to use this fairly abstract, fairly complex, rules

01:37.315 --> 01:40.765
that allow us to exchange in our day-to-day basis.

01:41.105 --> 01:43.485
So there have been a lot of work,

01:43.785 --> 01:48.605
toward this finding at which age infants can do this kind

01:48.665 --> 01:51.005
of very complex abstract processing.

01:51.505 --> 01:55.725
And it has also be a, a lot of debate in the field.

01:55.925 --> 01:59.005
I don't know if, you, you know, about this,

01:59.145 --> 02:00.485
man Noam Shamsky,

02:00.705 --> 02:03.765
but at some point he put forward the idea

02:04.195 --> 02:07.885
that our linguistic capacity was, biologically based

02:08.265 --> 02:11.725
and our capacity to manipulate, words

02:11.985 --> 02:15.965
and category of words like non verb determiner, ter

02:15.965 --> 02:17.325
and ter was innate.

02:18.105 --> 02:21.005
And there, there have been then a lot of debate

02:21.005 --> 02:23.365
because obviously if it's innate, it means

02:23.365 --> 02:26.485
that we can do it from when we are born.

02:27.145 --> 02:29.045
And there have been people that were not very

02:29.205 --> 02:30.245
happy with this idea.

02:30.665 --> 02:33.685
And so there have been a, a, a gigantic amount of data

02:34.265 --> 02:38.285
trying to find out if infants can manipulate syntax from

02:38.285 --> 02:39.445
fairly early on.

02:40.545 --> 02:45.045
And, my work, is a bit of a continuation of this.

02:45.545 --> 02:47.005
So while I wanted

02:47.005 --> 02:50.925
to know is at which age infants can do this manipulation?

02:51.105 --> 02:52.725
And one of the thing

02:52.725 --> 02:55.605
that has been difficult in the past few years for,

02:55.625 --> 02:57.965
developmentalists, for people working in the field

02:57.985 --> 03:02.605
of language development is to make the difference between

03:03.245 --> 03:06.605
a sentence that is grammatical.

03:07.145 --> 03:09.365
As I said, you will never produce some sentence

03:10.145 --> 03:12.165
and a sentence that,

03:13.375 --> 03:17.005
where the probability of words to appear together

03:17.585 --> 03:18.645
are super low.

03:19.345 --> 03:22.125
So one of the things that you, you need to understand is

03:22.125 --> 03:23.405
that we have a confirmed here.

03:24.115 --> 03:27.005
When I say that you are able to decode the sentence

03:27.005 --> 03:30.445
that you have never heard, it means that if I use correctly

03:31.095 --> 03:33.365
words that you, you you have never heard together,

03:33.945 --> 03:35.765
you will be able to decode even.

03:35.785 --> 03:38.885
So you have never been arbitrated to the sequence,

03:39.405 --> 03:40.525
specific sequence of the world.

03:41.275 --> 03:43.525
Obviously, it means that there is some,

03:44.035 --> 03:45.925
what we call transition of probability.

03:46.105 --> 03:50.125
So the probability that towards co-occur equal zero,

03:51.145 --> 03:54.685
and sometimes it's equals zero because it's ungrammatical.

03:54.685 --> 03:56.365
Sometimes it's equals zero

03:56.365 --> 04:00.965
because it's, it's just not used in your,

04:01.065 --> 04:02.165
in your input.

04:02.865 --> 04:05.125
And so this is a confirm that is very complicated

04:05.125 --> 04:08.685
to manipulate because it's possible that the infant react

04:08.745 --> 04:11.005
to the fact that there is a zero probability.

04:11.945 --> 04:15.045
And so it's not because it's grammatical, but it's

04:15.045 --> 04:17.765
because he has the, the, the, the children,

04:18.025 --> 04:19.605
the infants never heard the sentence

04:19.605 --> 04:22.245
before that he react,

04:23.065 --> 04:24.605
but not using the syntax.

04:24.625 --> 04:27.085
And it's very difficult to do this manipulation with,

04:27.085 --> 04:29.285
with infants because obviously they don't know a lot

04:29.285 --> 04:32.525
of words and they don't have a big span of attention.

04:32.905 --> 04:35.125
So there have been a lot of work around this.

04:35.785 --> 04:40.045
And, what we have decided to do, it was, a,

04:40.325 --> 04:43.205
a modalization so a feasibility experiment.

04:44.345 --> 04:48.445
And the goal here was to say, okay, so let's say that,

04:48.665 --> 04:51.285
early infants, no few words,

04:51.385 --> 04:53.925
he ha the infants has a she somewhere

04:54.155 --> 04:57.765
because, he, he has his, input from,

04:58.105 --> 04:59.245
you know, the caretakers.

04:59.705 --> 05:03.005
He has also the environment and some pragmatic ideas.

05:03.265 --> 05:06.845
So for example, if you show the bottle with the mic

05:07.075 --> 05:10.845
with the infants, this is an object that is fairly ate.

05:10.845 --> 05:13.765
Maybe at some point he will learn the meaning of bottle.

05:13.815 --> 05:17.365
Maybe he will also understand, the verb eat

05:17.365 --> 05:19.525
because he is eating sometimes

05:19.625 --> 05:23.085
and there is some correlation between the action

05:23.185 --> 05:24.205
and the, and the word.

05:25.465 --> 05:28.525
So we decided to do a, a, a, a model

05:29.155 --> 05:30.565
that will know some words

05:30.705 --> 05:34.945
and we wanted to see if there were this,

05:35.565 --> 05:38.185
this little amount of information

05:38.215 --> 05:42.105
that was bring from another part of our linguistic capacity,

05:42.105 --> 05:46.425
which is Symantec, how it can potentially help

05:47.165 --> 05:50.305
infant to start a syntactic calculation.

05:50.535 --> 05:52.345
What we did, again, it's a modernization.

05:52.405 --> 05:56.465
So what we wanted to know is how an infant will be able to,

05:57.065 --> 05:59.125
Uh, acquire New non

05:59.185 --> 06:02.445
and new verb, given the handful of non

06:02.505 --> 06:04.845
and verb that this infant will

06:04.845 --> 06:05.845
Have learned.

06:06.385 --> 06:08.565
So what we did is to,

06:08.865 --> 06:11.925
use a very helpful, source

06:11.945 --> 06:13.445
Of information. It's called

06:13.705 --> 06:14.705
Shy Death. So Child

06:14.705 --> 06:15.685
Death is a website

06:15.685 --> 06:17.045
That is open to everybody.

06:17.625 --> 06:18.925
And on this website

06:18.925 --> 06:20.445
There is Transcription of,

06:20.905 --> 06:23.605
Uh, Caretakers and infants, exchange.

06:23.985 --> 06:26.685
So basically scientists go to the home of people,

06:26.995 --> 06:28.285
they record the parents

06:28.345 --> 06:29.765
and their infants when they're together

06:29.825 --> 06:30.925
and their natural exchange,

06:31.225 --> 06:33.325
and we have the transcription. So

06:33.325 --> 06:36.085
What we did with that, it was to,

06:36.455 --> 06:37.725
First clean it a bit,

06:37.825 --> 06:41.645
and then we, selected the most often nouns

06:41.645 --> 06:44.125
and the most often verbs that the infants

06:44.715 --> 06:45.805
Have, heard.

06:46.265 --> 06:47.805
And we decided that those

06:47.805 --> 06:49.165
Words will be the words that the

06:49.355 --> 06:50.355
Infants. So,

06:50.355 --> 06:52.325
or model, no. So

06:52.325 --> 06:54.205
Basically that was the words that the,

06:54.475 --> 06:55.925
That the model Have,

06:56.105 --> 06:58.005
Uh, been, have been able to

06:58.115 --> 07:01.925
Acquire, just by The, the input.

07:03.295 --> 07:05.475
And what we decided is that our,

07:06.295 --> 07:09.435
our model will have some kind of semantic

07:09.805 --> 07:11.355
Basis. So

07:11.825 --> 07:13.395
What the model is doing is

07:13.395 --> 07:15.195
that it's separating from one side

07:15.255 --> 07:16.255
The verb.

07:16.255 --> 07:18.595
So the action, it's kind of an

07:18.595 --> 07:19.795
Approximation Because

07:20.085 --> 07:22.195
There is some known that, Uh,

07:22.305 --> 07:23.395
Also refer an action.

07:23.515 --> 07:24.715
I can say, for example, a

07:24.715 --> 07:27.155
Jump, A jump is an action, but it

07:27.155 --> 07:30.075
Is a known, but when you Look

07:30.295 --> 07:34.155
At the global vocabulary that we have, generally

07:34.295 --> 07:37.955
We can approximate that most of the action are used

07:38.015 --> 07:39.435
As Verb and most

07:39.575 --> 07:41.915
Of the objects are used, as

07:41.915 --> 07:43.315
Referring, to a norm.

07:43.735 --> 07:46.035
So we decided that basically the

07:46.105 --> 07:48.635
Lexicon of our modern, We have this

07:48.795 --> 07:50.675
Semantic information. It will,

07:51.205 --> 07:52.955
It'll know that,

07:53.455 --> 07:56.715
Few nos are object and few nos are action,

07:56.815 --> 07:59.075
and they are already a separation between

07:59.445 --> 08:01.515
Those Two element of the

08:01.545 --> 08:02.545
Lexicon.

08:02.865 --> 08:07.795
Once you have this, you can start to do a

08:08.035 --> 08:09.075
Distribution analysis.

08:09.575 --> 08:12.555
So what we decided was A very simple algorithm

08:12.855 --> 08:15.315
Is that The infants are able

08:15.495 --> 08:18.035
To look at all, Three grams.

08:18.175 --> 08:19.475
So three grams is

08:19.965 --> 08:24.235
Three wards, two, And learn all of them.

08:24.765 --> 08:27.515
Among the three grams that our model have learned,

08:27.845 --> 08:30.035
There is those seeds.

08:30.375 --> 08:32.795
So the nouns and the verbs that the model know.

08:33.335 --> 08:35.155
And what we have done is that we have

08:35.155 --> 08:37.755
Decided To run many model a model

08:37.905 --> 08:40.035
That was, very poor

08:40.055 --> 08:43.675
In lexicon because it has only four noun and two verbs,

08:44.255 --> 08:46.875
and we have double this amount each time

08:47.255 --> 08:50.515
to run different model until we have, until a,

08:50.555 --> 08:51.755
a model that is not,

08:52.375 --> 08:53.795
Uh, a, a very good

08:53.795 --> 08:55.325
Modernization of a, of a child

08:55.325 --> 08:58.885
because, the model knew all the non and all the verb,

08:59.105 --> 09:00.845
but it was just a way to see

09:01.155 --> 09:04.965
what this method could do if it was at the maximum capacity.

09:05.865 --> 09:09.485
And so we have look recorded all the sequence

09:09.485 --> 09:14.045
of three words, and then we have kind of request

09:14.065 --> 09:16.325
or model to do a categorization task.

09:16.865 --> 09:19.445
So we give a new part of this transcription,

09:19.785 --> 09:22.485
so a transcription that the model have not been trained on,

09:23.225 --> 09:27.485
and we ask the model to give the category of the third words

09:27.675 --> 09:32.645
that, the, that, the model encounter into this,

09:32.825 --> 09:33.885
new corpus.

09:35.225 --> 09:37.915
What the model is doing is that it's taking the two world

09:37.935 --> 09:42.035
of context before, guessing the third one.

09:42.055 --> 09:45.475
So for example, if the model encounter the little cat,

09:45.975 --> 09:50.355
it will, it'll have the, and little to guess the next word.

09:51.335 --> 09:54.395
And what we have done is that we have required a model

09:54.855 --> 09:59.035
to give exactly the, answer that is expected.

09:59.615 --> 10:02.515
So here you have to remember that a model has some kind

10:02.515 --> 10:03.995
of a mix flu

10:04.265 --> 10:08.115
because there is some words that,

10:08.615 --> 10:11.835
we expect a category that's the word that the model know.

10:12.095 --> 10:14.995
So we expect the model to answer non

10:15.055 --> 10:17.475
or verb all the words.

10:17.905 --> 10:21.195
They are just themselves that they sta as they are.

10:21.335 --> 10:24.155
So for example, the model doesn't know any adjective,

10:24.815 --> 10:29.035
so little skeletal, and there is some words non

10:29.055 --> 10:30.675
and verb that the model doesn't know.

10:30.735 --> 10:33.835
So again, they stay themself and we count true

10:34.335 --> 10:39.275
or correct only when the model gives exact category of the,

10:39.375 --> 10:40.475
of the words

10:40.475 --> 10:43.915
that in encounter the result were really surprising

10:43.915 --> 10:47.995
because I don't know if you realize how simple the model is,

10:48.135 --> 10:50.955
is just collecting sequence of three words.

10:51.815 --> 10:55.915
And what we have found is that the model is very precise.

10:56.295 --> 11:00.675
So the precision is when you say, for example, none

11:01.255 --> 11:03.475
how many times where you Correct.

11:04.845 --> 11:07.425
So the precision were, were excellent.

11:07.425 --> 11:11.865
It was around 80% for the non, a bit lower for the verb,

11:11.865 --> 11:14.185
around, 70% of the time.

11:14.185 --> 11:16.385
Correct. And what we have done

11:16.445 --> 11:18.705
to estimate if the model was good,

11:18.705 --> 11:21.425
because it's something that is difficult in development,

11:21.565 --> 11:23.065
we don't want a perfect model.

11:23.165 --> 11:24.705
We don't want a model that is able

11:24.705 --> 11:26.705
to categorize all the non and all the verb.

11:27.005 --> 11:30.425
We want a model that is able to model the infants

11:30.425 --> 11:32.625
and obviously the infants doesn't know all

11:32.625 --> 11:33.745
the non and all the verb.

11:34.125 --> 11:37.385
So then you have to estimate how good is your model.

11:37.445 --> 11:41.025
So what we have done is a very simple little manipulation.

11:41.485 --> 11:43.545
We have run a chance model.

11:43.965 --> 11:46.465
So a model that will just say non

11:46.465 --> 11:48.625
and verb adds the proportion of the non

11:48.685 --> 11:50.985
and verb of the training corpus.

11:51.405 --> 11:54.965
So it's a model that has no training compared

11:54.965 --> 11:57.125
to our other model that have learned something

11:57.435 --> 11:58.565
from the distribution.

11:59.385 --> 12:03.005
And, the, chance model was very below

12:03.385 --> 12:07.285
or the train model, especially on the precision.

12:07.285 --> 12:08.525
And then we use another measure

12:08.525 --> 12:10.405
that's a bit harder to understand.

12:10.475 --> 12:13.445
It's called recall, but, it's,

12:13.675 --> 12:15.245
it's not very hard to understand.

12:15.395 --> 12:18.365
It's basically when you need to find some things,

12:18.755 --> 12:20.285
were you able to find the things?

12:20.385 --> 12:24.365
So when, does the model have swe,

12:24.605 --> 12:26.725
a non at the level of the norm

12:26.725 --> 12:28.765
that was present in, in its target?

12:29.905 --> 12:32.285
And for this, it was also interesting

12:32.285 --> 12:36.445
because the, the recall was, a have a slop.

12:36.985 --> 12:40.925
So basically it was going up with a level of non

12:40.985 --> 12:44.165
and verb that the model knew from the beginning.

12:45.465 --> 12:48.205
Why is that? Uh, well, it's, or

12:48.205 --> 12:50.285
because we have given a mixed flu.

12:50.705 --> 12:54.445
So if we look at what the model was doing bad,

12:54.445 --> 12:56.805
especially in the low vocabulary,

12:57.115 --> 13:01.445
what the model was doing is that sometimes he encounter,

13:01.465 --> 13:06.285
it encounters a noun that, he guess was known.

13:06.785 --> 13:10.685
For example, after Al, Al is a context of noun,

13:10.805 --> 13:12.285
a very easy context of noun.

13:12.285 --> 13:14.165
So it was answering noun,

13:14.385 --> 13:15.845
but sometimes it was noun

13:15.845 --> 13:18.925
that the model didn't learn like elephant, a very,

13:19.225 --> 13:20.565
you know, uncommon nouns.

13:20.865 --> 13:23.885
But we expect that the model answer elephant here.

13:24.185 --> 13:26.685
So most of the time when we look at

13:26.685 --> 13:30.205
what the model was doing, it was not trying to mix non

13:30.205 --> 13:32.365
and verb, which will be very bad,

13:33.145 --> 13:37.525
but it was throwing an item instead of the category

13:37.865 --> 13:39.885
or the category instead of the item.

13:40.155 --> 13:43.285
With this model, we can see that the very fairly easy

13:44.445 --> 13:49.365
learning method can be used in fact to learn from the input.

13:49.495 --> 13:52.325
Where are your norms and where are your your verb.

13:53.305 --> 13:56.205
Uh, obviously this doesn't demonstrate

13:56.515 --> 13:59.045
that this is the method that the infants use

13:59.305 --> 14:00.765
to learn syntax.

14:01.665 --> 14:03.325
It could be useful to learn it,

14:03.785 --> 14:08.405
but we have made some, decision, some,

14:08.435 --> 14:10.725
arbitrary decision that we have to demonstrate

14:10.725 --> 14:14.845
that they are true to then be able to say, guy,

14:14.865 --> 14:17.045
we have found out how infants learn syntax.

14:17.585 --> 14:20.685
So in my future research, I want to,

14:21.065 --> 14:25.205
to study if it's, if this, semantic seed,

14:25.345 --> 14:30.125
if the way we have said, in fact, infants are able to say

14:30.125 --> 14:33.445
that the non are object and that the verb are action

14:33.545 --> 14:36.565
and those elements are very separating.

14:36.565 --> 14:40.605
The lexicon is true from the very first word that you learn.

14:42.205 --> 14:44.485
Interestingly, it's something that is true for adults.

14:44.585 --> 14:48.925
So if you do a MRI and a scan of a adult brain

14:49.345 --> 14:51.405
and you, present norms

14:51.405 --> 14:54.725
and your present ferb, you will see very fairly distinct

14:54.725 --> 14:56.045
region that will activate.

14:56.585 --> 14:58.805
So it's not a totally crazy idea,

14:59.105 --> 15:01.205
but it hasn't been checked in infants.

15:01.585 --> 15:04.685
Do they learn the kind of semantic that is attached

15:04.685 --> 15:05.725
to a world directly?

15:05.745 --> 15:08.685
And would it affect the way they anchor the

15:08.685 --> 15:09.885
world in the lexicon?

15:10.265 --> 15:13.925
We have to check that. We have to check also if once

15:14.555 --> 15:17.205
they have learned the distribution of the word,

15:17.595 --> 15:19.525
they can transmit this information.

15:19.585 --> 15:22.925
So let's say that the infants know that cat is an object

15:23.305 --> 15:25.805
and cat appear with oli.

15:26.385 --> 15:29.285
Are they able, when they encounter elephant

15:29.575 --> 15:33.285
after the little to say, in this case this object,

15:33.285 --> 15:35.325
this word should be an object?

15:36.355 --> 15:37.485
That is another question.

15:37.625 --> 15:39.965
And that's something that I would like also to test.

15:40.265 --> 15:44.845
So if it's true, it'll give a real cognitive, cognitive,

15:45.715 --> 15:49.405
cognitive assessment of, of, this model that was just

15:49.405 --> 15:50.445
for no affordability.

Exploring Syntax and Language Development in Infants

00:00:00 - When we use language, we, we don't really realize it because it's very automatic, for us adults, but we use many different units. And one of, the important one, and that has been argued to be, the facts that we are specific language is specific to human is, syntax. So syntax is our capacity to combine war together. And what is interesting is that we combine, words together according to rules that we are not totally aware.
00:00:33 - as a speaker of your own language. Some things like, flower on the, yellow, that doesn't make any sense. You cannot do this and you will never produce it. What is interesting also is that because you know, some of the rule of your language, you are able to integrate sentences that you have never heard before. So if I say something like, the cat is mm-hmm. Under the mm-Hmm.
00:01:05 - -mm are you start to understand what is happening. You have a fairly good idea of the action that the cat is doing and where is doing it. So, this capacity seems to be very specific to human. It's called generative. So we can generate and decode sentences that we have never heard. And so one of the question that is crucial for my work is to know from when infants can do that, when do we start
00:01:37 - that allow us to exchange in our day-to-day basis. So there have been a lot of work, toward this finding at which age infants can do this kind of very complex abstract processing. And it has also be a, a lot of debate in the field. I don't know if, you, you know, about this, man Noam Shamsky, but at some point he put forward the idea
00:02:08 - and our capacity to manipulate, words and category of words like non verb determiner, ter and ter was innate. And there, there have been then a lot of debate because obviously if it's innate, it means that we can do it from when we are born. And there have been people that were not very happy with this idea. And so there have been a, a, a gigantic amount of data
00:02:38 - fairly early on. And, my work, is a bit of a continuation of this. So while I wanted to know is at which age infants can do this manipulation? And one of the thing that has been difficult in the past few years for, developmentalists, for people working in the field of language development is to make the difference between a sentence that is grammatical.

Decoding Syntax: Infants and Language Learning

00:03:10 - and a sentence that, where the probability of words to appear together are super low. So one of the things that you, you need to understand is that we have a confirmed here. When I say that you are able to decode the sentence that you have never heard, it means that if I use correctly words that you, you you have never heard together, you will be able to decode even. So you have never been arbitrated to the sequence,
00:03:41 - Obviously, it means that there is some, what we call transition of probability. So the probability that towards co-occur equal zero, and sometimes it's equals zero because it's ungrammatical. Sometimes it's equals zero because it's, it's just not used in your, in your input. And so this is a confirm that is very complicated to manipulate because it's possible that the infant react to the fact that there is a zero probability.
00:04:15 - because he has the, the, the, the children, the infants never heard the sentence before that he react, but not using the syntax. And it's very difficult to do this manipulation with, with infants because obviously they don't know a lot of words and they don't have a big span of attention. So there have been a lot of work around this. And, what we have decided to do, it was, a, a modalization so a feasibility experiment.
00:04:48 - early infants, no few words, he ha the infants has a she somewhere because, he, he has his, input from, you know, the caretakers. He has also the environment and some pragmatic ideas. So for example, if you show the bottle with the mic with the infants, this is an object that is fairly ate. Maybe at some point he will learn the meaning of bottle. Maybe he will also understand, the verb eat
00:05:19 - and there is some correlation between the action and the, and the word. So we decided to do a, a, a, a model that will know some words and we wanted to see if there were this, this little amount of information that was bring from another part of our linguistic capacity, which is Symantec, how it can potentially help infant to start a syntactic calculation.
00:05:52 - So what we wanted to know is how an infant will be able to, acquire New non and new verb, given the handful of non and verb that this infant will Have learned. So what we did is to, use a very helpful, source Of information. It's called Shy Death. So Child Death is a website That is open to everybody. And on this website There is Transcription of,

Infant Language Acquisition Through Semantic Analysis

00:06:23 - So basically scientists go to the home of people, they record the parents and their infants when they're together and their natural exchange, and we have the transcription. So What we did with that, it was to, First clean it a bit, and then we, selected the most often nouns and the most often verbs that the infants Have, heard. And we decided that those Words will be the words that the Infants. So, or model, no. So
00:06:54 - That the model Have, been, have been able to Acquire, just by The, the input. And what we decided is that our, our model will have some kind of semantic Basis. So What the model is doing is that it's separating from one side The verb. So the action, it's kind of an Approximation Because There is some known that, Also refer an action.
00:07:24 - Jump, A jump is an action, but it Is a known, but when you Look At the global vocabulary that we have, generally We can approximate that most of the action are used As Verb and most Of the objects are used, as Referring, to a norm. So we decided that basically the Lexicon of our modern, We have this Semantic information. It will, It'll know that, Few nos are object and few nos are action,
00:07:59 - Those Two element of the Lexicon. Once you have this, you can start to do a Distribution analysis. So what we decided was A very simple algorithm Is that The infants are able To look at all, Three grams. So three grams is Three wards, two, And learn all of them. Among the three grams that our model have learned,
00:08:30 - So the nouns and the verbs that the model know. And what we have done is that we have Decided To run many model a model That was, very poor In lexicon because it has only four noun and two verbs, and we have double this amount each time to run different model until we have, until a, a model that is not, a, a very good Modernization of a, of a child because, the model knew all the non and all the verb,
00:09:01 - what this method could do if it was at the maximum capacity. And so we have look recorded all the sequence of three words, and then we have kind of request or model to do a categorization task. So we give a new part of this transcription, so a transcription that the model have not been trained on, and we ask the model to give the category of the third words that, the, that, the model encounter into this,

Modeling Language Learning Through Contextual Guesses

00:09:35 - What the model is doing is that it's taking the two world of context before, guessing the third one. So for example, if the model encounter the little cat, it will, it'll have the, and little to guess the next word. And what we have done is that we have required a model to give exactly the, answer that is expected. So here you have to remember that a model has some kind of a mix flu
00:10:08 - we expect a category that's the word that the model know. So we expect the model to answer non or verb all the words. They are just themselves that they sta as they are. So for example, the model doesn't know any adjective, so little skeletal, and there is some words non and verb that the model doesn't know. So again, they stay themself and we count true
00:10:39 - of the words that in encounter the result were really surprising because I don't know if you realize how simple the model is, is just collecting sequence of three words. And what we have found is that the model is very precise. So the precision is when you say, for example, none how many times where you Correct.
00:11:07 - It was around 80% for the non, a bit lower for the verb, around, 70% of the time. Correct. And what we have done to estimate if the model was good, because it's something that is difficult in development, we don't want a perfect model. We don't want a model that is able to categorize all the non and all the verb. We want a model that is able to model the infants and obviously the infants doesn't know all the non and all the verb. So then you have to estimate how good is your model.
00:11:41 - We have run a chance model. So a model that will just say non and verb adds the proportion of the non and verb of the training corpus. So it's a model that has no training compared to our other model that have learned something from the distribution. And, the, chance model was very below or the train model, especially on the precision. And then we use another measure
00:12:10 - It's called recall, but, it's, it's not very hard to understand. It's basically when you need to find some things, were you able to find the things? So when, does the model have swe, a non at the level of the norm that was present in, in its target? And for this, it was also interesting because the, the recall was, a have a slop. So basically it was going up with a level of non

Modeling Infant Syntax Acquisition Challenges

00:12:45 - Why is that? well, it's, or because we have given a mixed flu. So if we look at what the model was doing bad, especially in the low vocabulary, what the model was doing is that sometimes he encounter, it encounters a noun that, he guess was known. For example, after Al, Al is a context of noun, a very easy context of noun.
00:13:14 - but sometimes it was noun that the model didn't learn like elephant, a very, you know, uncommon nouns. But we expect that the model answer elephant here. So most of the time when we look at what the model was doing, it was not trying to mix non and verb, which will be very bad, but it was throwing an item instead of the category or the category instead of the item. With this model, we can see that the very fairly easy
00:13:49 - Where are your norms and where are your your verb. obviously this doesn't demonstrate that this is the method that the infants use to learn syntax. It could be useful to learn it, but we have made some, decision, some, arbitrary decision that we have to demonstrate that they are true to then be able to say, guy,
00:14:17 - So in my future research, I want to, to study if it's, if this, semantic seed, if the way we have said, in fact, infants are able to say that the non are object and that the verb are action and those elements are very separating. The lexicon is true from the very first word that you learn. Interestingly, it's something that is true for adults. So if you do a MRI and a scan of a adult brain
00:14:51 - and your present ferb, you will see very fairly distinct region that will activate. So it's not a totally crazy idea, but it hasn't been checked in infants. Do they learn the kind of semantic that is attached to a world directly? And would it affect the way they anchor the world in the lexicon? We have to check that. We have to check also if once they have learned the distribution of the word, they can transmit this information.
00:15:49 - and cat appear with oli. Are they able, when they encounter elephant after the little to say, in this case this object, this word should be an object? That is another question. And that's something that I would like also to test. So if it's true, it'll give a real cognitive, cognitive, cognitive assessment of, of, this model that was just for no affordability.

Faculti AI can make mistakes

Research

Policy

Analysis