Notes

we are the paradoxical ape bipedal naked
large-brained long the master of fire
tools and language but still trying to
understand ourselves aware that death is
inevitable yet filled with optimism we
grow up slowly we hand down knowledge we
empathize and deceive
we shape the future from our shared
understanding of the past
Carta brings together experts from
diverse disciplines to exchange insights
on who we are and how we got here an
exploration made possible by the
generosity of humans like you

thanks very much and thanks all of you
for coming it's wonderful to be here and
talk to such a incredible audience I
thought I'd start off by posing a
question that I think is particularly
relevant to all of us
thinking about language evolution and
that's what's so special about language
and in a sense the goal of evolutionary
linguistics is to try and explain the
origins of this special properties of
language and there are two that I want
to focus on today in fact Roger started
with one of them so first off I want to
highlight structure so language has a
highly unusual and rich systematic
structure in the way it's put together
so when I say the phrase to boots you
are able to relate the acoustic waveform
that hits your ears to this rather cute
picture of my son's shoes on this slide
now how are you able to do that they
were able to do that because of the
special structure that language has the
signals that we produce are constructed
of sequences of these meaningless
elements we call them phonemes so these
different phonemes are put together in
different different ways reorganized in
different ways to make up the sounds of
our language in sequences but
furthermore these meaningless elements
are themselves put together to make
morphemes the meaning carrying elements
of language and these are put together
to construct meaningful wholes
so you understand what to Bhoots means
because you understand these ways in
which language recombines elements and
it's because of this recombination
because we can endlessly recombine
elements at these two different levels
that we are able to talk about anything
so we're able to constantly be producing
completely
you utter insist happy expectation that
everyone can understand this and this is
an incredible feat and remarkably it
seems to be unique to our species okay
so that was one of the special features
of language structure so next up the
other one I want to talk about is
learning so the way we are able to use
language because we have learned how
both how to produce the signals in our
language and we've also learned perhaps
even more remarkably what the signals in
our language mean so we are able to use
language meaningfully because we've
observed language being used around us
as we as we grow up again this this
combination of learning the signals and
learning the meaning seems to be unique
in nature what I'm particularly
interested in is that the fact that this
is a very particular kind of learning
when we learn language we're learning
from the product of other people who
have gone through the same learning
process so we we learn from the
observation of the output of other
learners so this is something that I
call iterated learning and it's a
particular and relatively unusual
process in nature
so what iterated learning means is that
the language I I speak I learned by
observing language being produced by my
parents and the language I produce it's
going on to effect the language that my
son and daughter are acquiring and so
what this means is that we have this
process of cultural transmission of
language through this mapping from
external production of language and an
internal some internal representation of
language and what this means is that
languages evolve so this is a kind of
cultural evolution so the language is
language itself is a an evolutionary
system in its own right and what I want
to suggest today that these two features
structure and iterated learning are
related to each other and in fact what I
want to claim is that language structure
is the inevitable
or consequence of the fact that language
is a culturally evolving system so I
want to try and pose and answer this
question affirmatively that cultural
evolution can explain why language has
the structure that it does so how else
could we go about answering this
question so this is one of the
challenges of doing evolutionary
linguistics is how on earth do you study
it scientifically well there are lots of
ways you're going to hear about a lot of
them today but one of the ways that we
pioneered in my lab in Edinburgh was to
try and recreate language evolution the
cultural evolution of language in the
experiment lab so I'm going to show you
how we do that I'm gonna give you two
examples of two experiments that we've
done that will give you a flavor of the
way in which we can study cultural
evolution of language in the lab so
generally how do we do this so what we
did is we brought together two to kind
of standard experimental techniques from
two different areas one was a paradigm
from psycho linguistics called
artificial language learning where we
get participants to into the lab and
teach them miniature languages and then
test them and another technique from
experimental anthropology I guess you'd
call it called a transmission chain
paradigm and this is a way people study
the process of cultural evolution in
various different domains so we're just
plugging those two together so I'll talk
you through in general how it works and
then I'll give you there and two
specific examples so first off we bring
a participant into the lab and we ask
them to learn a miniature language and
so this is a very very small miniature
artificial language that we've created
so we might have them in the lab for an
hour maximum and then we test them so we
get them to produce utterances in this
miniature language so at this stage this
is a standard paradigm artificial
language learning paradigm but the twist
we put on it is that we use the output
of that participant when we test them to
form the language that the next
participant in the lab is going to learn
from and then we take the output of that
second participant and use that to form
the input training data for the third
participant in the lab and so on so we
create a chain of transmission of this
miniature language and watch how it
evolves as it passes from participant to
participant in our experiments what we
typically do is we start with language
in scare quotes that is random and
unstructured and doesn't have these
structural properties that we're
interested in and then we want to see if
those structural properties emerge
through this process of transmission
over these artificial generations that
we create and in the lab experiments ok
so that's the general structure of these
these experimental that approach and I'm
going to give you two examples so here's
the first example this is joint work
with Hannah Cornish and Kenny Smith
pictured there and this was this is our
first experiment the first experiment we
did in our lab and so this is as you can
see a 2008 it's a relatively recent
approach that we've been taking here and
what we wanted to do is look for this
emergence of compositional structure so
compositional structure is the idea that
and you can put words together to make
meanings so the meaning of a whole
expression is made up of the meanings of
parts of that expression so we wanted to
see if we could get that to emerge in in
these lab experiments okay so what
participants did so we have it this kind
of experimental chain of participants
and what they participants were asked to
do was learn a miniature language and
their language was made up of these
weird meanings which were just colored
shapes that were moving in different
ways and we had three colors three
shapes and three different types of
movement so 3 times 3 times 3 that makes
27 different possible meanings in this
language so when I say miniature
language I really mean miniature okay
and these pictures these meanings were
paired with strings of syllables and
these were just made up at random by the
computer
okay so this is completely unstructured
arbitrary language every single meaning
was given a different label and each
participant was trained on half of the
language but then was tested on all of
the meanings so they were they were
trained on half of all of half of the
meanings but then we were asked to
produce strengths of syllables for all
of the meanings when we tested them and
then their output then formed the input
the next participant in the lab and
again we took a random half of the
output of the first participant tutoring
the second one and then test the second
one and take another random half of
their language to train the third
participant and so on down the chain
okay so what happens well what we find
is that the people are absolutely
terrible at this task so so the this
graph shows error and it's rather
forgiving measure of error it just said
an error of one mint would mean that
you've got not a single character not a
single letter right in the target word
and error zero means you get everything
right so and what you see here is four
different runs of the experiment and
each point on this graph is a different
participant okay so the first
participants in the lab had a really
terrible time but nobody got a single
word exactly right so it's court
essentially absolutely awful by the end
however the participants in the in the
later generations in these chains some
of them were getting everything exactly
right including the meanings that they
weren't trained on so they were
successfully guessing the correct label
for meanings that we didn't even test
that they're trained them on okay so how
on earth is that happening well I'll
show you what the languages look like as
they evolve so here's here's an initial
random language so for example in this
table the word mini key is the word for
a Blue Square that's moving horizontally
and now remain paramount if you're
trained on half of this language I'm
actually if I blanked out half of these
words and then
asked you to reconstruct the missing
words that you'd have no hope you
couldn't possibly get it right okay
because this is random so if you didn't
know that that word was mini key there's
no way you could guess it so that's why
they're doing so badly and this is what
it that same language looks like ten
generations later right so now you see
how the participants are doing this
right so now if I blanked off half of
those words you'd be able to
successfully guess what the missing
words were for example the word poin a
refers to all spiraling objects so this
language has evolved to become more easy
to learn but it's evolved in a very very
kind of slightly disappointing way from
our point of view so it's a it's a bold
just by jettisoning words in a
particular way so that these words refer
to two particular sort of sections of
the set of meanings so why was this
happening in fact in some versions the
experiment would go down to a single
word for all they're all things
so clearly we need something else in
this experiment so we need some kind of
pressure to be expressive as well as
learnable for the language I can't go
into the details we don't have time but
what we essentially did was we added in
a filter on the output of each
participant so we threw out items that
were ambiguous before we passed it on to
the next participant and none of the
participants could be aware that we were
doing this manipulation so this is so we
ran the whole experiment again with this
extra step hidden step and we got
completely different results so here's
again the initial and another initial
language again completely random and now
and now ten generations later it looks
quite different now what I've done here
is I've added in hyphens to make it
easier for you to see but the
participants don't get to see these
hyphens right and what you see here is
that different parts of the signals
correspond to different parts of the
meanings so for example this n prefix
means the black things the r prefix
means red things a close suffix means
something that's bouncing and so on so
this is the compositionality we were
looking for and it emerged out of this
process of cultural transmission so
structure seems to emerge spontaneously
from this process of cultural evolution
now there's a couple of obvious worries
that you might have so firstly perhaps
the participants in our experiment are
kind of deliberately thinking will this
this language is rubbish
here's how we can fix it up we'll add
some structure well you don't think
that's what's happening because bear in
mind these two different runs of the
experiment got very different languages
out but the participants couldn't know
which which condition they were in
another possible objection is well okay
but they already speak a language in
fact all of our participants spoke
English so it's not surprising that they
gradually change this system to be one
that looks more like a language well
that's a valid worry so we've recreated
that all of this experiment in computer
simulation and where we can know that
our
computational participants don't have a
pre-existing language and we get the
same results okay so that's my first
example the second example and we want
it to move away from this starting point
with these random strings because that
seems a bit weird where'd they come from
and instead we start from a potentially
much more natural starting point and one
of gestural pantomime and we wanted to
see if we could evolve structured
something looked like a structured sign
language miniature a structured sign
language out of initial state that was
iconic pantomime and this is joint work
with Kenny Smith catio barometer and
Eric Urquhart mill so here's how the
experiment works so in this version of
the experiment the meanings now are
videos of a bouncing ball so here's one
so that's that's the things that have to
be conveyed and the signals are manual
gestures made to a video camera so
something like this and the initial
language in this experiment was
improvised one-off pantomimes that we
got people into the lab and just said
okay can you give the pantomime that
corresponds to this video participants
are trained on twelve out of the sixteen
videos but asked to produce gestures for
all sixteen and then the same process of
this sort of diffusion chain
transmission chain was applied so we'd
see how the gestures evolved over
generations I'll just show you the full
set of meanings it's a little bit
alarming when you see it so these are
all the different videos piled together
and what you can see here is there is
different paths in which the ball moves
along so flat sloping in an S shape or
in a circle and there are different
manners so the ball might be bouncing or
rolling or spinning or giving this
nightmarish jittering motion so that's
the set of meanings they have to convey
and this is what the initial pantomimes
look like for all the participants we
got in just to do one-off gesture for
these and what you see here is there's a
huge range of different strategies that
people use to produce these meanings
there's an enormous diversity in
strategies here so what we were
interested in yeah some of them more
successful than others what we are
interested in is how do these strategies
evolve down generations so here's our
first participant who saw 12 of these
and then produced a gesture for each of
them now it looks more systematic but
that's because it's one person so that
they will look a little bit more similar
in fact there's still a lot of diversity
here if you look at the hand shaped used
to convey the ball and so on there's
still a lot of diversity much later in
the same chain we have this participant
now it's looking much more systematic
there's there's if you look at a
handshake for example it's now very
similar across all the different videos
and there's particular kinds of
systematicity we were interested in so
here's here's an example from another
chain which I particularly like and have
a look at what she's doing here to
convey the ball moving so she signals
the ball and then says the manner and
then the path so she separated out two
aspects of this meaning compositionally
into two different parts and indeed
another interesting thing that she does
here she's self-corrects so she says oh
no I got it wrong and watch again she'll
do it again
and this indicates that at this point in
this the evolution of the language he's
seeing that there's a right and wrong
answer it's become grammatical okay so
just to summarize these and then I'll be
done so over time the gestures in this
language become less pantomimic and more
conventional and we can actually measure
this we can show that systematic
structure emerges over generations in
this experiment and interestingly this
separation of manner and paths mirrors
what we can see in real emerging sign
languages so this is work by a nice a
house and colleague colleagues looking
at how in Nicaraguan sign language and
emerging sign language we get the
separation of manner and path emerging
over generations over cohorts
so just to find final conclusions and I
want to argue that language is an
evolutionary system in its own right
and languages because of this language
is adapt and they adapt to pass more
easily from learner to learner and
linguistics structure is the solution
that cultural evolution finds to the
problem of being learner both so it's an
inevitable consequence of the way
language is transmitted and we can and
indeed we should study this process in
the experiment lab thank you very much