Irene Eleta: Multilingual Users of Twitter: Social Ties Across Language Borders or How a Story Could Travel the World

Good morning, everyone.

Thank you for coming here
[unclear] of the semester.

So, I'm going to start.

Access to the internet
is greater than ever before

and as a consequence,
it's becoming more multilingual.

However, there's evidence of segmentation
of cyberspace

due to language and national borders.

This image serves to illustrate that.

This is the language communities
of Twitter in Europe.

So, what you can see are tweets
geolocated over a map of Europe

and the different colors
represent the different languages.

You can even see regional languages
like Catalan in the Catalan region of Spain

And this is going to be useful
for an example I'm going to use later.

I'm interested in Twitter in particular,

because of the speed
of information dissemination

and that most of this information
is publicly accessible.

I'm going to illustrate this
with a capture

of a dynamic visualization
you can find on the Twitter blog

by Miguel Rios.

And what you can see here
is the global flow of tweets

after the earthquake in Japan.

In pink, there are the tweets
coming out of Japan

and, in green, the retweets
all over the world.

This illustrates that in Twitter
information is spreading across countries.

But how can this happen?

Expatriates, migrants, minorities.
diaspora communities, language learners

all play an important role
in building transnational networks

and cultural bridges
between nations and communities.

They are the multilingual users
on the internet.

The overarching research question is:

how are multilingual users of Twitter
connecting different language groups?

In 2009, the Berkman Center of Internet
and Society at Harvard University

mapped the Arabic blogosphere

and they described a key concept
for my research.

They discovered an English bridge
and a French bridge of bloggers

that were writing in their native
Arabic language and in English or French.

And they were connecting the different
national blogospheres

with the international one.

This might have played a role in the Arab
popular uprisings in 2011

for reaching out to the world.

And this is connected with a concept
that first appeared in 2008

of the bridge bloggers.

So, bridge bloggers are bloggers

that are trying to connect
their local communities

to a wider global audience.

The image you can see here
is actually the visualization they created

of mapping the Arabic blogosphere.

Each dot is a blogger, or a blog.

The size represents their popularity,
so how many incoming links they have

and they grouped them--

the neighborhoods they created

in relation to the linking
between them.

So, the ones that are grouped together
are linking among each other.

The colors are a different question.

The colors represent "attentive clusters",
that's how they call it.

And they look at their online resources
and media outlets

these blogs were linking to.

So, blogs of the same colors
are following the same media outlets

and online resources.

And they did human coding
to label those groups.

And here is where we see
the label English grids

the responses from Cuba
in English

and up there, there's [unclear] France.

And so I think it's important to retain
the concept of attentive clusters.

Now, let's go back to 2011
during the Arab popular uprisings.

And I'll show you a visualization
of the influence network

of Twitter users in Egypt.

So, what you're seeing here

just imagine people down the street
at Tahrir Square

tweeting in Arabic about what's going on
on the ground.

And those are the people in red.

So, these red dots represent users
that are tweeting in Arabic.

Then we have the international community

or even Americans, British and so on
tweeting in English.

And they are in blue,
those blue dots.

And then, interestingly, we have
people in between them.

which are illustrated in different
degrees of violet, or violet shades.

This represents the fact that they
are tweeting in both Arabic and English.

So, what we're seeing
is the bridge Twitters

because, like Ethan Zuckermann called them
"bridge bloggers".

So, another context.

The same year, 2011, a lot
of big protests were going on in Europe.

And in particular, in Spain.

They started on May 15th 2011
there were massive protests.

And of because of this context,
this situation

new attentive clusters were emerging
in the social media landscape of Spain.

Now, this is a visualization you can find
in the Socialflow blog, research blog

on social networks.

And what it is, is it tracks the origin
and the initial spread

of the hashtag #occupywallstreet
in Twitter.

They detected that one of the first users
of the hashtag #occupywallstreet

was on July 13th 2011, linking to a blog
post of Adbusters.

So you have the Twitter account
of Adbusters there, very big

because it's being retweeted a lot.

And mentioned a lot.

And they collected these mentions
and the tweets that had these mentions

and these retweets with the hashtag
during July 13th.

From July 13th to July 23rd.

So, from the first 10 days
of the use of this hashtag

it was from the very beginning of the use
of this hashtag on Twitter.

They just mapped the accounts

and the series of posts with the hashtag
and mentions with the hashtag

and the users that were connecting

because of these mentions
and retweets.

Now the interesting thing
in this visualization

is that they

the Socialflow people
particularly in [inaudible]

detected this Spanish brand
of users

were forming an attentive cluster.

Mentioning and retweeting about it
in Spanish

using the hashtag in their messages
in Spanish.

And they point out in the blog

that this Spanish contingent

helped post and spread the word
about Occupy Wall Street

even before most of the United States
was aware of it.

So, I found that very interesting.

And it was due to the context
in Spain at that moment

with big protests and new clusters
forming in the social media landscape.

Now I have shown you the importance
of these multilingual users

in connecting language communities
and spreading information

across countries, acting as mediators.

But let's focus on another aspect
of connecting language groups

which is language choice.

So I'm going to devote a moment
to speak about languages

and language choice.

To understand languages in the world

I'm going to use a telescope.

So de Swaan...

...proposed a theory called
the world language system

back in the 1990s.

to explain the languages in the world.

And he used a very beautiful metaphor,
the constellation.

So, in his theory there's about a dozen
languages in the world

that are the hearts of the system,
or the suns.

The suns of the system.

For instance, English, French, Spanish,
Arabic and more.

And then there are hundreds,
maybe more than 100, 200...

national languages that are orbiting
around these suns like planets.

And finally we have regional
and minority languages

that are orbiting these planets
like satellites.

And he used this metaphor
to explain the power relationships

between languages.

This is a theory of what he called

"communication potential
and language competition"

A key point he made

is that the system holds together

thanks to multilingual people
and interpreters.

This is what's providing cohesion
to the system.

He also made a controversial proposal

about the communication potential
of a language.

So, he proposed a formula,
a mathematical formula

where he could estimate the communication
potential of a language

and supposedly a person with tools
through learning and usage

based on the communications of that.

For example, a person might decide
to learn English and use English

because not only does it provide
communication with English native speakers

but also, adding to that, it provides
the possibility to communicate

with all the second-language learners
of English

from many different languages,
many different countries.

So, supposedly, in history

English provides
the greatest communication.

And he received some criticism,
because of the central role of English

in his theory

He said it was the central hub
of all the system.

There's also the language ecology paradigm
first proposed by Haugen in 1972

and there's this idea of an ecosystem
of languages

and, again, it's using another metaphor

and because of this metaphor

also appeared the idea
of endangered languages.

I'm going to briefly just read
the definition.

He defined the language ecology as:

"the study of interactions between
any given language and its environment"

and what I think is very important:

"language exists only in the minds
of its users"

which leads me to point at my research.

In my research, I'm using a microscope
to see the cells

and my cells in my study
are the Twitter users.

Why is that?

Because as Haugen explains,
there's a psychological dimension

to language ecology

where language interacts
with other languages

in the minds of multilingual people.

And there's a sociological dimension
to language ecology

where we use language to communicate
and interact with other people.

And this language ecology generates
because of the people

that decide to use that language

learning and interacting
with people using it.

And this is the point
of language choice in languages.

So, I focus on the connections of people
and the language choice.

So, these are the four points
I'm going to be speaking about.

But actually the main focus
is going to be the first point

Social network analysis and the taxonomy
of intersections between language groups

This is where I'm going to be spending
most of the time.

And then very briefly,
just for compilation purposes

I'm going to speak about another
small study that I did

the factor analysis

looking at the influence
of the social network

in the language choices of the users.

So, how the social network
influences language choice

of our multilingual users.

And then I'm going to briefly also talk
about the last study of my dissertation

that is still ongoing.

So, I still have new research
to talk about.

And it's content analysis

and in this case I'm focusing on
intrinsic factors

intrinsic to the messages

about the topic,
and the type of exchange.

If it's a reply,
if it's a public post

and how that influences
the language choice as well.

And finally I will...

I'm going to give you my reflections

so I can invite your thoughts
and suggestions and discussions about it.

Briefly, I'm going to start
with the sampling

so I can talk about the rest
of the research.

So my focus is on multilingual users,

how did I identify multilingual users
on Twitter?

It was giving me a headache.

Finally what we decided...

this research has been--

I have always had the help
of Jennifer Golbeck,

she was my adviser.

And I did this with her help.

So what we did, was gather a list
of what is called stopwords.

From different languages
and you have a list over there.

And then the stopword lists
you can find them on the internet.

They are created
for computational linguistics

so they use it for filtering purposes.

And they are common words
in a language.

Very common words in a language.

So, sometimes they're used precisely
for eliminating them from texts

when they're in, for example,
searches in Google

the eliminate the stopwords,
the stopwords that you type

in the search.

But in this case I wanted
to find the stopwords

that are very common in the language
to represent the language.

And so we had to select words
that were not written the same

as in another language.

Sometimes, could be confusing
and ambiguous.

Then I typed in Google...

one word in one language
and one word in another language.

Usually I was always using
one English word

and one word in a different language.

And I looked in the Twitter domain.

So the search results from Google
will give me the profiles

of people on Twitter that in theory
wrote messages in both languages.

We had to do a lot of hand-combing
to actually see if it was in two languages

or it was just that they were mentioning
an English song

the title of an English song
but they had no English in the rest.

So we had to ensure
that they were authoring tweets

in two languages.

So writing them, not just retweeting them

they were not just automatic postings
from Facebook.

So we had a long set of criteria
a lot of manual combing

and then finally we selected
92 multilingual users

and in total they used 19 languages,
2 or 3 languages per person.

Now, I don't know if you want to ask
some questions about the sampling

because there's a lot of details about it.

No doubts?

Or maybe they'll come later!

Now, how do I do
the social networks analysis?

Well, now I have my 92 multilingual users
technically they are called the ego

of an egocentric network.

This is the cell of my study.

It started with the nucleus of the cell

which is my multilingual user

and then I go to Twitter

and first of all I have instructed--

so in this case my ego
is called the Painter

and I have extracted the last 50 messages
that he posted on Twitter

to see the languages
this person used-- is using.

And I see that he is using English,
Spanish and Catalan.

Catalan is a regional language in Spain

and I have shown you on the map
the region before

where the region was.

And they speak both Catalan
and Spanish.

So, this person is tweeting
in a minority language

a national language

and also international.

So, I already found the Painter
and I know what languages this person speaks

well, uses on Twitter,

and then I extract
all the social networks.

So, the followers on Twitter

you know that on Twitter
you have followers

and you follow people.

I extracted both.

The followers of the Painter
the people that are following him on Twitter

and also how the friends
are connecting to each other.

So, all of them, all of these dots
are the followers

the people following the Painter
on Twitter

and also I see how they connect
among each other, ok?

So the Painter follows Eduard
in the center

and it seems he's very popular.

And then I extract the last 30 posts
of Eduard--

there's a reason for that

but vernacular
is mostly economy questions!

I will tell you why!

So I extracted the last 30 posts of Eduard

and then I do
automatic language identification

with the Google API
for language identification

which costs money.

So you have to really think
about how many posts you want to send

to Google and how much money
you have available

and what is the accuracy
you're going to have

according to how many posts you send.

There's a lot of testing going on there.

I do the same with everybody
in the social network.

I extract the last 30 posts

use the Google identification

build that algorithm that decides
based on the languages of these 30 posts

is this person monolingual?
Is this person multilingual?

Which languages?

And then I laddered them, ok.

This is just a visualization behind the--

Perhaps person 1 is monolingual,
or bilingual of two languages.

Now that I have all the friends
of the Painter

how they connect,

I color code them
depending on the languages they are using.

And here, what you can see
is very interesting.

I don't know if you can distinguish
the colors well

because up here, this area,
that is like a triangle

there's a group of users
writing in English.

And it's pink.
Sort of pinkish.

And then, down here
there's this Spanish group

in light green.

And, in the middle, the one
that perhaps doesn't distinguish as well

from the English,
is the Catalan group.

So the users writing in Catalan
in dark blue.

And then there's a set of violets
in between

and these violets represent
the bilingual users

either English and Catalan
or English and Spanish.

And then there's darker green
around here,

they are using both Catalan and Spanish.

So there's a lot of bilinguals
going on.

And there's an interesting dynamics

in that you have this English group
up there

and the Spanish group up here

and the Catalan group in the middle.

And this Catalan group is very mixed up
with the Spanish group

which makes sense,
because it's a bilingual community.

So, this is how I built the egocentric
network of my 92 multilingual users.

The Painter is just one of them.
I have 92.

I have 92 cells or egocentric networks
that I studied with my microscope.

Do you want to ask some questions
about this process

or this visualization?

(person 1) Of the bilingual units,
are they users or tweets?

They are users, yeah.

So, the dots represent people.

So, like Eduard here.
They represent people.

Now each dot to determine the language
and the color

I extracted 30 posts

So, it's an interesting question
because the 30 posts

have different language levels
assigned to them

especially if they were bilingual

and I had to decide which language level
I was going to assign to the user.

So, I had to build an algorithm
with a set of rules

basically saying--

the Google identification system
would give me a language

and a confidence level

So if the confidence level was very low

I would say "discard that"
because I had a series of pluristics

based on both the number of tweets
using a particular language

and also on the confidence level.

And there are a lot
of technical challenges there as well.

(woman) So, it's possible
that some of these posts

many of these posts would be multilingual,
I'm sorry monolingual in one language or the other?

So it's also possible that some
of these individual posts

would mix languages?

Yes, it is possible.
It's very possible!

It's very challenging
for the automatic system!

(woman) Right, ok.
I just wanted to be clear--

Yes, exactly.

So it's not as frequent as I expected,
having bilingual posts

that I would call.

But it's happening.

And so, for a series of tests,
I had to do manual combing

and I saw that sometimes
it was the case

that they were doing some sort
of translation in the same tweet

and sometimes it was just the case
that they were mentioning titles of things

or places in a different language.

So, there's a lot of issues
surrounding the automatic handling of this

but you are dealing with 92 networks

and they have between 30
and 5,000 nodes in them.

So, I don't remember the numbers exactly,

but I'm talking about
around 80,000 people.

So detecting the language of 80,000 people
and this is small-scale.

If you go to millions,
you need an automatic system.

And one of the things I'm having
to write up in my dissertation

is what are the challenges.

You have to be prepared for them,
to solve those problems.

And one of them is what do you do
with bilingual posts

which language do you assign to that post?

Automatic posts, spam...
there's a lot of problems.

Challenges, I mean.

That's what makes it interesting
because you cannot do manual combing

on these scales.

Do you have another question?

So, now, what am I doing with this?

I'm going to classify my social networks,
looking at the patterns

of overlaps between the languages groups.

And overlaps or intersections.

I'm looking specifically at the networks
that have only two language groups

I had five of these networks
that were trilingual

so I put them aside to go simple
first with just two language groups

to see how they interconnect.

And then I classified them

first following a qualitative analysis

and then I used network statistics
that I developed with my adviser

for this purpose.

And I will talk later a little more
about it.

So, tried to provide
more robust measures for that.

I classified them and I came up
with some types.

This is what I call the gatekeeper
language bridge type.

And there's some variants of it,
obviously.

What you can see here
is the network of a person

and I'm going to assume this person
is in the United States

and speaks both Spanish and English.

Let's call her Maria.

So she's Maria and she has two groups
of friends using Spanish on Twitter

and then that big group of friends
using English.

And, as you can see,
there's just a few nodes

connecting the two language groups.

You can see that the social structure
can be different from the language groups

so you can have maybe a group of friends
and a group of coworkers

inside the same language group,
so it can be more complex

than just dividing the social network
by language groups.

There can be more grouping
because of other social resources.

But the interesting thing is that
there are only a few nodes

where people are connecting
holding together these Twitters.

I think this was friends
with English here.

You can see, in this case, it seems
like the two groups

are holding closely together

because there are much more links
holding the two groups together.

Of course, this is going to depend
on the size of the networks

so I had to account for the size
when coming up with measures

with network connections

I had to provide ratios.

Now, the ratio of [close] language linking
here and here

and you have these types--

These types are not just clear-cut.

There's an evolution.

There's people that have
very few connections

with the language groups

and then progressively there's people
with more and more.

And this increases.

Which points to the fact,
that my cells are there.

Which means I don't see the evolution
over time, ok?

This is a limitation of my research.

I just see the social network
of this person looked

at a particular point in time.

I don't know how it evolves over time.

So, for myself, it's just there.

It would be interesting
to see these different patterns

that I have been observing.

Maybe over time these connections
between languages maybe increasing.

Now we have the integration
and union type

where in this case you have a person
from an Arab country

and green represents the friends
that are using Arabic

and the friends using English are in pink,
but there's also violet

there are bilinguals.

That means there's a group
of English users

and bilingual English - Arabic users
inserted in the group of Arabic, inside.

That's the integration,
so they're integrated.

And then I have a Greek guy,
who uses Greek and English

and his Arabic friends.

And in this case, you can see
it's sort of light blue

representing Greek, so the friends
that tweet in Greek

Pink again represents people tweeting
in English

and there's a lot of bilinguals.

So these kind of dark blues
represent the bilinguals.

And these are two groups

that if you've seen before,
the gatekeeper and the language bridge

progressively getting closer and closer

with more and more links
across languages.

In this case, this is like the extreme.

The links between the two languages
are so dense

that you cannot almost distinguish
where the border is

between the two language groups.

And, interestingly, the border might be
even only noticeable

because there's a lot of bilinguals
around it.

And this is the union type
where they unite.

And finally, the peripheral language type.

This is a Brazilian guy,
the network of a Brazilian guy

where you have--

probably he lives in the United States
or something like that--

because this guy has mostly
all this big group of friends

tweeting in English.

And then there's the side tentacle
running outside, using Portuguese.

And this is like a periphery landscape.

So, in the periphery there's a small group
of Portuguese language.

Now, I forgot to mention that there's dots
that are light yellow or white.

Those are the ones that have no data.

So, I don't know
the language they're using

because either their accounts are closed

or for some reason, in between the collection
of data they closed the account.

Mostly, the reason
is that they're private accounts

where you cannot get the data from.

I think somewhere I read
it was about 5 percent.

I'm not sure.

But for one reason or another,
I don't have that information.

Now, why am I classifying them?
These networks?

Well, the reason is that--

well, there are some studies
that demonstrate that the social structure

the structure of the social networks
influences the spread of information.

How information disseminates
in the network.

So, I'm just assuming
that these different structures

are going to influence the spread
of information.

But this is a study that has to be done.

I cannot demonstrate that one
of these types

facilitates the spread of information.

I can only say that I am assuming,
so that potential study

could just look at, for example,
if gatekeeper and language bridges

are not as good for spreading information
as union and integration types.

Right, we can just assume
because of the cross-language links

so, how many links there are
or the ratio of discourse language

may potentially facilitate information
diffusion in these cases.

So, that study needs to be done.

I cannot say what's going to happen!

I just assume it's going to be like that.

So that is the reason why I classify them.

I have some network statistics.

We've made about an 80 percent accuracy
guess, which is quite good,

but the sample is small.

So now, do you have any more questions
before I move past to the next study?

man) I was curious as to how many--

what was the selection process like
to find the 92 users?

Well, this is what I've been spending
the beginning

about just using two stopwords
from two different languages

typing that in the search box in Google
and searching Twitter

and then once--

Basically you just go through
the list of results

and start opening the profile,
counting the tweets.

How many in this language,
how many in the other.

And we put a threshold of 10 percent

they had to have written 10 percent
of the tweets in a second language

and you couldn't count retweets
or automatic posting.

We also had to manually discard
these spammers.

So, that was the process.

(woman) And that's a paid search
through Google?

No, that we did manually

and then once--

So the other thing you can say is you can
use these core multilingual users

and then do what I did for behavior
in these social networks

which is once you extract the friends
and extract the messages of the friends

and automatically find the language

then you can say "Oh, this person
is multilingual" automatically.

You just process it and you can detect
a lot more multilingual people

through that process.

The paid process was sending these posts

to the Google language
identification tool.

So, what I did was clean each message
automatically.

Basically, eliminating the hashtags

and the mentions
that had an @ in front,

symbols, URLs, all those things
I would automatically eliminate them

and then with the rest of the message,
I'd send that to the Google API

for language identification

and the Google API would give me
a language level and a confidence binary.

And that for each message.

And then I built the algorithm
with the help of Jen Golbeck

to decide, well I have 30 messages,
500 English

10 million Spanish and then one in Swahili
which is unlikely

and you had to decide
the confidence value--

So I used rules, defined rules

but it could be done
statistically I think.

And write some statistical method
to decide

"well this person actually is bilingual"
or whatever.

That's the process.

It's long!

Yes.

(woman) Hi, I understand
that you did it manually

but currently in existing research field
is there any software

that we can use to capture,

to have access to all
these different tweets?

And to capture the different categories?
[inaudible]

Ok, so you mean the extraction?

(woman) Yeah.

No, I didn't do it manually.

(woman) And the other,
I think the other part

of your data presentation
is visualizations coming out

like this graph.

Can you show us what kind of research
do we have for social scientists

to present the data in a visual form?

This is a tool I would recommend.

[inaudible]

So, the first question.

All the extraction from Twitter,
it was automatic.

I didn't copy the tweets,
it was automatic.

I used the Twitter API.

They have a process
for registered developers

and I extracted it automatically.

Now, the tools, and I forgot
to put that in this slide

but in the beginning,
when I showed you the first visualization

I put the name of the tool in--

I don't know if I translate well,
but I think it's G-E--

You can see here, G-E-P-H-I,
I don't know how to pronounce it!

["Jefy" I think...]

So, this is the one I've used
for the visualizations

and it's good because you can use it
on any platform.

So both on a Mac or a PC or Linux.

Now, it has limitations for...

mostly for network statistics
in my opinion.

The other one, that is very popular
is Node XL.

And in fact it was developed
here in the ATI lab.

In the lab where I work.

So, they collaborated with Microsoft.

It's a template for Excel

and it allows--

In fact they are still adding new features
and there's two people working on it

in the lab.

But the reason I haven't used it here,
is because I have a Mac

and also there's another reason
I like this positioning algorithm

and this is...

this is another issue
I haven't talked about

is how you actually place the dots.

And actually these algorithms for layout
use force-directed schemes

like in physics science.

So if a node has a lot of links
with another node

they put it closer,
so it's like there's forces

or strings attaching the nodes.

And depending on how many strings
there are, they're closer or farther.

There's physics science rules
for placing them.

But there's different algorithms

but the other reason I chose Gephi
is that it has an algorithm

specifically in this tool
that places my language groups separately

more than any other algorithm
that I could use in Node XL.

And it was more useful
to see the groups separated.

But you can use both
depending on what you want to do.

They both have weaknesses and strengths,

different depending
on what you have to do.

Node XL has more features
for processing many networks

and extracting network statistics
for many networks at the same time.

And it has a lot of interesting features,
maybe this is more manual.

I don't know.

Somebody called it
"the Photoshop of visualization".

So I'm going to briefly comment
on the factor analysis.

The point here, what I want to see
is multilingual users of Twitter

are aware of their audience in a way.

And they somehow perceive
how many followers

of this language or the other they have.

Maybe not very consciously,

but they perceive something.

So, I went to see how this social network

the fact that there's many languages
or just one in the social network

can affect the choice of language in this person,
the ego person.

So, I actually did a lot of testing,
different variables,

but I'm just going to focus
on the essence,

which is I have my dependent variable
which is the proportion of English

used by the ego has 50 posts,
maybe 60 percent of them are in English

and 40 percent in Spanish,
I don't know.

And then they have the factor
of how many users in the network

are in English
and how many are using other languages.

And then the multilingual index
of the network

- and this is my favorite part -

because it's basically saying

"is multilingualism encouraging English
as a lingua franca?"

especially on Twitter, where we have these
public posts that anybody can read.

So anyway... I'm not going to go
into the technical details

of bi-nodal statistical interpretation.

What I wanted to do is
that in these combined effects

of the factors,
which one was more important?

Was heavier than the others?

Had more weight in defining these
proportional [inaudible] used by the ego.

I tried other factors,

I also looked at the use
of non-English language

In the end... there are certain,

I mean, they're obvious somehow.

I think it's more interesting the process
of what I've learned

than the results themselves.

Because basically what I've learned
is that, yeah,

the English use of the network

is encouraged by the use
of English by the ego

and in a certain way it's so important
that any other factor

is really not that important.

And even the second most important,
the multilingual index

was so light compared with
the heavy impact of English

used in the network.

But what I thought was really interesting

was how do you define
the multlinguality of a network?

And with this I got help
from Jordan Boyd-Graber

who is also in the iSchool

and in the lab for computational lab,
the information processing lab

here in Maryland.

He helped me
with all these technical aspects.

And he was the one suggesting
"Well, why don't you look--"

"instead of just looking at the number
of languages in the network...

"because sometimes you get
wrongly detected languages...

like Swahili. Well, no one was really
speaking Swahihi in this network.

There were technical challenges,
like I explained to you.

So maybe there's a high number
of languages in the network

but the network is mostly monolingual.

Mostly everybody uses English
and just a few people maybe use others

or maybe just it got wrongly detected.

And maybe you're just saying

"Oh yeah, there's ten languages
in the network!"

and actually it's not
a very multilingual network at all.

So, we came up with this, the entropy.

And this is a physics concept
that measures the disorder

in a system.

And in this case, the entropy
would be my multilingual index

and what it's doing is providing a value
between 0 and 1

So, with 0 it's a very homogeneous system
everyone speaks the same language

and if it's closer to 1,
it's really a heterogeneous

and it places an importance

in how many people
are using its language.

So, this is the equation,
just to show you it.

And it takes into account the number
of languages in the network

and then one of the variables
is how many nodes in that language

that there are divided by the total number

and this is what gives the proportion
for example.

So just to let you know
that there's interesting lessons

from this study.

Despite the research not being exciting!

And this is what I'm doing right now.

So, the intrinsic characteristic
of the message

how that influences the language choice.

First, I'm wondering,
because I just saw it in the content

are replies encouraging people
to use their native language?

And public posts encouraging people
to use English as a lingua franca?

This is one that showed up the same.

And I changed the handle,
for privacy reasons...

So this is the reply to somebody
and it's in Arabic.

And this is a public posting
and it's in English.

Now, the thing I'm looking at
is public analysis

and I'm considering with Jordan
to do some automatic topic analysis

because there's many languages,
so I cannot decode it all

in many of them.

Only in three, maybe four...

So, I'm wondering,

are technology topics favoring
the use of English?

And other topics,
international news maybe?

Whereas other topics
like national news or songs

they might be encouraging the use
of native languages.

And then I'm looking
if there's translations

or if there's cross-cultural words
that you can detect.

For instance, this person
is writing in English

but it recommending a visit to a museum
in the city of Lille in France.

So this person knows the city in France,
knows that to visit the museum

you go there.

And this is what I call
cross-cultural words.

[What I kind of found] is that surprisingly
there's not many translation behaviors

going on, despite these people
being multilingual.

And this is what is going to trigger
some reflections.

How am I doing on time?

(woman) 1:22.

(man) Umm, it's usually an hour long...

So, I will go on with my reflections.

to encourage some thoughts.

So the greatest connecting power
is the will of users who want

to be connected.

This is a really nice quality,
because the communities of interest

in social media, in Twitter
is what is bringing people

from different countries, together.

And also experiences,
like the Voluntweeters,

so after the earthquake in Haiti,
there were these spontaneous

self-organizations of Twitter users
for translating tweets

and they called themselves Voluntweeters,
there's a paper about that--

So this is the triggering
of social connections

across countries, across borders
and across languages.

But even when the social structure
could potentially facilitate

information diffusion
and cross-language linking

this condition is not sufficient.

There are other factors
like the design of the interfaces

and the design of systems
that can influence...

can promote, or not translation behaviors
and cross-cultural awareness.

And the Wikipedia
of cross-language linking

you have links for many languages
for every article.

We also still acknowledge the dynamic
language preferences of multilingual users

so they could address their messages
to the appropriate audience.

I like the solution of Google+
with their circles

where I can put my friends and family
in Spain in a circle

and write them in Spanish.

And then the recommendation of people
based on language profile

would be useful for this spontaneous
self-organization.

So, these are some of the things.

The impact of mediation.

Global Voices is
an international community of bloggers

that connect bloggers and citizens
from around the world

in different languages.

And Scott Hale

a student from Oxford University
led a very interesting study

after the earthquake in Haiti about blogs
in Spanish, Japanese and English

and he looked
at the cross-language linking

and focusing on this topic
over time.

And he discovered that 50 percent
of the cross-language linking

was happening through this platform,
Global Voices.

So, it had a very big impact
in the language links.

And finally, social media,
big media outlets,

people are interconnected
in these complex networks

and underlying is this language ecosystem.

So we have the language ecosystem,
and on top of that

we have the social media ecosystem.

People would share a video from YouTube
on Twitter, or news on Facebook.

What happened if we integrate
in this ecosystem

these platforms, like Global Voices,
like Universal Subtitles

which is a platform
for crowdsourcing subtitling of videos

and translation of subtitles
for videos.

If you integrate that and this
starts connecting, starts building paths

between languages,
that didn't exist before.

So I think we should make it easy
for multilingual people to translate

and subtitle all the content they like,
their favorite content

and share it with the appropriate audience
so they can start connecting

the language islands of the internet.

And that way stories will travel
all over the world.

Particularly I would like to thank
Jen Golbeck, my adviser

and Fulbright for supporting
this research.

And then I open the space
for questions and your ideas

if this has triggered some thoughts.

(woman) I have a question
about how this relates

to your Yahoo award.

Well, they have the Internet Experiences
lab in California.

And they--

So, we tend to think
maybe it's a super tiny place

but actually there are fields

and I applied for the social systems.

The social systems are a category.

And I think that was embedded
in the Internet Experience lab

and yeah, they liked it.

(man) But is it this
work that they are interested in?

Yes.

- The languages?
- Yes.

Well, now I have results,
because I wrote up reports

about what my work was about.

Great.

Yes?

(woman) I was thinking about
if you analyzed the place...

like if there's any relationship
between tweeters and tweets

and the place that the people are.

I mean, because it's not the same
being a Brazilian in Brazil

and tweeting in Portuguese
or being Brazilian in the US

and tweeting in Portuguese--

There's many, many factors
that I haven't looked at.

It's not part of your study?

But because I had to scope it somehow.

There's so many factors.

Geography was one that I was originally
intending to look at

but I found there were so many problems
to actually get the right geography

the right geolocation.

The problem is that I didn't originally
collect the geolocation.

I think only a small percentage
of messages have...

geolocated information.

I'm not sure about the percentage there.

So there's only a small percentage
of messages that have geolocation.

There's issues with the accuracy...

What I have collected is the information
in their profile

they can put the information
about the place,

but sometimes it's more
or less trustworthy,

sometimes there's nothing,
and sometimes there's just crazy stuff.

(audience laughs)

So, something absolutely has to be there.

If I wanted to expand this,
geography would be a nice place to go!

(woman) Ok.

Yes?

(man) Could you say a little bit more

I think you said about the visualization
choices you made?

Oh yes, well...

I tried this tool, the Node XL,

I used both Node XL and Gephi.

There's more...

I think there's, I don't remember the name
there's one that was developed

here in Maryland

but it's not as user-friendly.

But I've forgotten the name,
I will have to look it up.

And there's a lot of tools
that are for really technical people

that are handling millions of nodes.

Because with these tools,
for social scientists or humanists

maybe they are not.

Some tools can have maybe 300-400 nodes
and still be understandable.

But if you go beyond that,
actually visualizations get crazy

and even for more technical tools
for more technical people

there are hundreds or millions,
they cannot do visualizations

at some point they just give you
statistical measures.

I have to leave it out.

I have a list of tools and that

but if I need the names,
I need to go through everything.

(woman) But yours was Mac-accessible?

Yes, this Gephi tool is Mac-accessible,
you can use it with Microsoft

with Mac and with Linux.

And I forgot to say,
it's open source.

(woman) Did you find
studying languages and internet

was like a place, unexplored?

Like here in the United States?

Like when you began studying
or analyzing this

you felt that a lot of people
are doing this

or nobody is doing this

and I'm the first one trying to--

I'm not the first one,
but it's a very new area

to be exploring.

So, it's very exciting
because of that.

Because there's so many
unanswered questions

and I find that surprisingly enough
the United States is not paying so much attention

about multilinguality issues

And I think that language policies
are very monolingual-oriented

but it's terrible

because there's a whole lot
of multilinguality in this country.

There's so many people
speaking different languages

that I'm so amazed
about that contradiction.

Because in Europe,
it's an obvious challenge for us

because we need to understand each other
between all these countries

of the European Union.

And there's a lot of money invested
in research that relates to multilinguality

and communication in languages

and technology in particular,
cross-language systems

and in libraries there's a lot of work
going on.

There's investment in the research.

So yeah, maybe in terms of investment

the European Union is
not a bad place to be.

Better than the United States!

But at the same time,
what I find interesting

is that here when I talk about it
people are really interested

and interested in the subject
and excited about it.

Maybe in Europe it looks more
like old news.

Like "yeah, we already know that."

(audience laughs)

So I find that it's exciting
to be seeing the audience

like "Oh yeah!"
It's so new.

*(woman) Yes.

(woman) As the emerging view
of research in the United States

can you show me which institutions
or which area of academic institutions

actually have more invested
in this topic in the US?

I'm not sure about the institutions.

What I know, particularly,
in Indiana there's work

because Susan Herring
is a researcher there.

She has inspired my work.

She published a book
The Multilingual Internet

and she has done research on blogs,
also communities

of different languages connecting blogs
in the blogosphere.

So she has been one of the ones,
one of the first tackling these issues

and she's still going
and she's doing something.

So, it's the University of Indiana,
I think.

Yeah, Susan Herring.
Look for her!

And also at the same university
there's Paolillo.

He's also doing research
in this area

and he actually published for UNESCO
for research on language diversity

on the internet.

So Susan Herring and Paolillo,

they are at the same university.

Those are my inspiring ones.

Well, at Harvard at the Berkman Center
of Internet and Society also did

this mapping of the blogs.

But they don't focus on languages.

But there's tangential thing
around there.

(man) One more question?

Well, thank you very much!

Thanks!

(audience applauds)