thank you very much yeah so we were
talking about quests early alright so
making the internet safer places you
know as good a place as any to start I
think and so yes we’re not allowed
basically we tried to do stuff around
risk safety security trying to make the
internet a safer place well I’m going to
talk to you about today specifically is
hate speech and cyber hate so how many
of you here use online social networks
about seventy five percent how many of
you would know how to engage with cyber
hate if you came across it to try and
kind of remedy it about four people
great so I got a job I would have been a
short talk otherwise so I’m going to
talk to you a little bit today about how
we define it how we measure it how we
observe it then given that we are you
know research scientists and how we can
manage it and the real takeaway for you
guys hopefully will be that will be the
management now first thing I’m going to
do is distinguish between hate crime and
cyber hate so hate crime is enshrined in
law so in the UK we have the Public
Order Act which basically has strict
thresholds which needs to be passed for
something to be a crime that is based on
protected characteristics race religion
sexual orientation and so on but for
something to be a crime it actually
needs to have either an incitement
element to it or an active threat now
that to us doesn’t seem all that
suitable for managing something like the
internet it’s too big too many people on
it and there’s too much stuff out there
at the moment as I’m sure you’ve seen so
our take on this is that if we’re going
to measure something we need to measure
something that can be understood by
everyone so we take a harm’s perspective
basically if somebody can be harmed by
it then it’s cyber hate I saw our take
on it
now these things tend to be triggered by
things like brexit the Westminster
attack president Trump we’re going to
build a wall Mexico is going to pay for
it right these things tend to lead to a
reaction an online reaction an online
reaction targeting people we protected
characteristics following the 9/11
attacks the offline hate crimes recorded
58 percent of those happen within two
weeks so we’ve got to be able to react
to this quickly and this is offline
so imagine online we need to be able to
measure this quickly and in real time
okay so what we’re essentially looking
for is to measure antagonistic and
hateful language online now you all put
your hands up views and online social
networks most people now use online
social networks now when you’re using
them for good which everyone is in it
today right you’re all promoting TEDx
and all the speakers that are here today
yep so you’re using it for good and
trying to get a positive message out
there it gives you this megaphone right
gives you this stage this platform to be
able to reach everyone anywhere in the
world that’s got access to online social
networks which is now a large proportion
but unfortunately it also gives a
platform to people with views that are
hateful
and our antagonistic and they want to
express and they want to react to things
you know several people today have
mentioned about how people react to
things and people come back people get
stressed things like changes in politics
and government you know they tend to
lead to that so we get these reactions
now this also then leads to echo
chambers and when cyber hate is out
there people tend to join in with this
and the longer it’s out there for and
the more it gets propagated through
these online social networks it remains
a problem so our job is to measure this
better understand how to manage that
type of problem now all online social
networks suffer with this but we’ve
particularly studied Twitter because
it’s so open it is so highly connected
what we’ve done is basically collected
data from Twitter following one of these
trigger events in particular this was
Woollett and the and the attack on Lee
Rigby and Willett
and basically what we’ve done is we’ve
asked people is this hateful
antagonistic simple as that where people
agree we take that as then a gold
standard this is hateful antagonistic
now one of the problems we face
following you know trigger events is
that there’s such a volume of reaction
in online social media we couldn’t give
each tweet out to every individual it
would take forever to annotate so what
we can do is train machines to recognize
this and recognize the features of cyber
hate and this is precisely what we’ve
done now how many of you have come
across machine learning before a handful
basically machine learning is
recognizing the features that we as
humans would look for to be able to
identify something so if I give you a
basic example of apples and bananas
right one feature we might identify as
the color so we might train a machine to
say if something is red or green it’s an
Apple the sentence yellow is a banana
there will be occasions of course where
you’ll have green bananas and green
apples and then you get confused too
I’ll have other features shape for
example if it’s round it’s an apple long
and thin it’s a banana right so machine
learning is basically identifying
features that a machine can recognize
and classify as something in our case
cyber hate so one of the obvious things
we could have done is to look for
derogatory terms I’m sure you can all
think of derogatory terms to two people
you could exhibit that hoot towards
people who exhibit to particulate
protected characteristics which fireless
actually doesn’t work all that well the
reason being is on inspecting some of
these tweets that we had annotated we
actually identified there was a much
more nuanced antagonism in here I’m sure
you’ve all come across it it’s this
message of S&M not here go home not
welcome okay now that doesn’t always
work with words themselves so what we’ve
actually done is to start look at
looking at probabilistic language models
which I won’t go into but essentially
what they’re looking for is co-occurring
terms like go back not welcome go home s
and then and they look for the ones that
probabilistically likely to occur
in cyber hate as opposed to general
everyday language so what we found is
that that using this method allows us to
be around about 80% accurate at
classifying something as cyber hate heat
full or antagonistic language which
means that we can train the machine to
basically do what a human would do
because let’s remember humans are 100%
accurate either during this annotation
exercise there was normally about 75% at
the time that people agreed as to
whether Turman is hateful antagonistic
so now that we can actually do this we
could do it at scale so following an
event we can take all the data that
comes in around the event and study it
so following the attack in Willets this
is precisely what we did what we wanted
to know was how viral does hate speech
go and we measured statistically how
likely different elements different
features of the tweets the statistical
power of these elements in predicting
whether they’d be retweeted or not
so we had things like whether they
contain a hashtag whether they contain a
URL who they were tweeted by we looked
at different media media outlets
political groups celebrities number of
followers one of the elements in the
model was cyber hate another element in
the model was sentiment positive or
negative fairly you know rough-and-ready
type of measurement the cyber hate we
know was about 80% accurate we
hypothesized the following this attack
cyber hate would be highly likely to be
retweeted you would imagine it would it
was actually 45% less likely to be
retweeted than a tweet that didn’t
contain cyber hate so it was not
spreading following this attack that was
supported by the sentiment measurement
that actually told us that it was 24%
more likely to be retweeted if it
contained positive sentiment this was
messages of support messages for the
public leaders that came out and
denounced the attack cyber hate didn’t
spread following this attack
we also looked at the longevity how long
does something stay out there for in the
Twittersphere this is important because
the longer it’s out there longer it’s a
risk that people will join on the
bandwagon create these echo chambers
cyber hate died out within a day we
measured a two-week period it died out
very very quickly didn’t stay out there
for very long so contrary to what we I
guess popular belief on online social
networks that they’re kind of echo
chambers of hate and antagonism we
measured it we’ve seen it it’s there
absolutely it’s there we can build
networks around looking at who
propagates that but it doesn’t spread
far and it doesn’t spread long which is
a good thing so this brings me to my
final bit I guess which is it’s just how
to manage this now all the other stuff
I’ve spoken to about so far is kind of
solid we you know we’re quite confident
in this there’s this other stuff now
we’re starting to get into we’re
starting to get some findings around how
to manage it now what we did is we
looked at threads on Twitter so if you
post a tweet and people replied to it
and they reply to it and they reply to
it and then applied to it and these can
get quite long so we looked following a
cyber hate remark how long these threads
would get and we measured again
statistically what were the factors that
meant they’d end up as long threads or
short threads the idea being can we see
how it stemmed now what we found was if
you go in there and call someone a
keyboard warrior absolutely lights a
fire the heat gets worse it gets more
extreme it gets really really very much
longer drawn-out we go in there and call
someone a racist actually they tend to
stop validation they’ve got what they
came for but what we also saw was that
if this happens in small numbers kind of
creates a bit of a hook for the person
who’s posted this original message to
come back you know someone’s kind of I
maybe validated me but someone else has
said someone else I can hook on there’s
really not much point engaging in this
at all
and the
the most useful remarks were just cause
to order you can’t do that it’s not
acceptable or requesting evidence and
there isn’t any evidence but the most
important finding is that the more
people individual people who get
involved in counter speech in this
thread trying to stop it trying to give
a message that says this is unacceptable
the shorter the thread became so
essentially what we’re seeing is
sensible online community activism so
following the riots in London right this
is an offline event but people came out
with their brushes and said look you
know we’re still here we’re going to
clean up we’re going to sort this out
you know you’re not going to beat us
this is in the offline world but the
online in the offline world can’t be
separated any more right when people are
attacked online they feel it in their
person and they feel it just in their
homes they don’t want to go out the
house they feel it when they go to work
they feel like everyone’s looking at
them just because something is online
doesn’t mean it doesn’t affect them in
the offline world we live in an online
in an offline world social cohesion is a
very big problem at the moment and cyber
hate online is there and is available
for anyone in the world to see the web
connects us all so my final remark will
be is you’re all online online social
media so if you see it work together and
you can stop it thank you very much
[Applause]
talking about quests early alright so
making the internet safer places you
know as good a place as any to start I
think and so yes we’re not allowed
basically we tried to do stuff around
risk safety security trying to make the
internet a safer place well I’m going to
talk to you about today specifically is
hate speech and cyber hate so how many
of you here use online social networks
about seventy five percent how many of
you would know how to engage with cyber
hate if you came across it to try and
kind of remedy it about four people
great so I got a job I would have been a
short talk otherwise so I’m going to
talk to you a little bit today about how
we define it how we measure it how we
observe it then given that we are you
know research scientists and how we can
manage it and the real takeaway for you
guys hopefully will be that will be the
management now first thing I’m going to
do is distinguish between hate crime and
cyber hate so hate crime is enshrined in
law so in the UK we have the Public
Order Act which basically has strict
thresholds which needs to be passed for
something to be a crime that is based on
protected characteristics race religion
sexual orientation and so on but for
something to be a crime it actually
needs to have either an incitement
element to it or an active threat now
that to us doesn’t seem all that
suitable for managing something like the
internet it’s too big too many people on
it and there’s too much stuff out there
at the moment as I’m sure you’ve seen so
our take on this is that if we’re going
to measure something we need to measure
something that can be understood by
everyone so we take a harm’s perspective
basically if somebody can be harmed by
it then it’s cyber hate I saw our take
on it
now these things tend to be triggered by
things like brexit the Westminster
attack president Trump we’re going to
build a wall Mexico is going to pay for
it right these things tend to lead to a
reaction an online reaction an online
reaction targeting people we protected
characteristics following the 9/11
attacks the offline hate crimes recorded
58 percent of those happen within two
weeks so we’ve got to be able to react
to this quickly and this is offline
so imagine online we need to be able to
measure this quickly and in real time
okay so what we’re essentially looking
for is to measure antagonistic and
hateful language online now you all put
your hands up views and online social
networks most people now use online
social networks now when you’re using
them for good which everyone is in it
today right you’re all promoting TEDx
and all the speakers that are here today
yep so you’re using it for good and
trying to get a positive message out
there it gives you this megaphone right
gives you this stage this platform to be
able to reach everyone anywhere in the
world that’s got access to online social
networks which is now a large proportion
but unfortunately it also gives a
platform to people with views that are
hateful
and our antagonistic and they want to
express and they want to react to things
you know several people today have
mentioned about how people react to
things and people come back people get
stressed things like changes in politics
and government you know they tend to
lead to that so we get these reactions
now this also then leads to echo
chambers and when cyber hate is out
there people tend to join in with this
and the longer it’s out there for and
the more it gets propagated through
these online social networks it remains
a problem so our job is to measure this
better understand how to manage that
type of problem now all online social
networks suffer with this but we’ve
particularly studied Twitter because
it’s so open it is so highly connected
what we’ve done is basically collected
data from Twitter following one of these
trigger events in particular this was
Woollett and the and the attack on Lee
Rigby and Willett
and basically what we’ve done is we’ve
asked people is this hateful
antagonistic simple as that where people
agree we take that as then a gold
standard this is hateful antagonistic
now one of the problems we face
following you know trigger events is
that there’s such a volume of reaction
in online social media we couldn’t give
each tweet out to every individual it
would take forever to annotate so what
we can do is train machines to recognize
this and recognize the features of cyber
hate and this is precisely what we’ve
done now how many of you have come
across machine learning before a handful
basically machine learning is
recognizing the features that we as
humans would look for to be able to
identify something so if I give you a
basic example of apples and bananas
right one feature we might identify as
the color so we might train a machine to
say if something is red or green it’s an
Apple the sentence yellow is a banana
there will be occasions of course where
you’ll have green bananas and green
apples and then you get confused too
I’ll have other features shape for
example if it’s round it’s an apple long
and thin it’s a banana right so machine
learning is basically identifying
features that a machine can recognize
and classify as something in our case
cyber hate so one of the obvious things
we could have done is to look for
derogatory terms I’m sure you can all
think of derogatory terms to two people
you could exhibit that hoot towards
people who exhibit to particulate
protected characteristics which fireless
actually doesn’t work all that well the
reason being is on inspecting some of
these tweets that we had annotated we
actually identified there was a much
more nuanced antagonism in here I’m sure
you’ve all come across it it’s this
message of S&M not here go home not
welcome okay now that doesn’t always
work with words themselves so what we’ve
actually done is to start look at
looking at probabilistic language models
which I won’t go into but essentially
what they’re looking for is co-occurring
terms like go back not welcome go home s
and then and they look for the ones that
probabilistically likely to occur
in cyber hate as opposed to general
everyday language so what we found is
that that using this method allows us to
be around about 80% accurate at
classifying something as cyber hate heat
full or antagonistic language which
means that we can train the machine to
basically do what a human would do
because let’s remember humans are 100%
accurate either during this annotation
exercise there was normally about 75% at
the time that people agreed as to
whether Turman is hateful antagonistic
so now that we can actually do this we
could do it at scale so following an
event we can take all the data that
comes in around the event and study it
so following the attack in Willets this
is precisely what we did what we wanted
to know was how viral does hate speech
go and we measured statistically how
likely different elements different
features of the tweets the statistical
power of these elements in predicting
whether they’d be retweeted or not
so we had things like whether they
contain a hashtag whether they contain a
URL who they were tweeted by we looked
at different media media outlets
political groups celebrities number of
followers one of the elements in the
model was cyber hate another element in
the model was sentiment positive or
negative fairly you know rough-and-ready
type of measurement the cyber hate we
know was about 80% accurate we
hypothesized the following this attack
cyber hate would be highly likely to be
retweeted you would imagine it would it
was actually 45% less likely to be
retweeted than a tweet that didn’t
contain cyber hate so it was not
spreading following this attack that was
supported by the sentiment measurement
that actually told us that it was 24%
more likely to be retweeted if it
contained positive sentiment this was
messages of support messages for the
public leaders that came out and
denounced the attack cyber hate didn’t
spread following this attack
we also looked at the longevity how long
does something stay out there for in the
Twittersphere this is important because
the longer it’s out there longer it’s a
risk that people will join on the
bandwagon create these echo chambers
cyber hate died out within a day we
measured a two-week period it died out
very very quickly didn’t stay out there
for very long so contrary to what we I
guess popular belief on online social
networks that they’re kind of echo
chambers of hate and antagonism we
measured it we’ve seen it it’s there
absolutely it’s there we can build
networks around looking at who
propagates that but it doesn’t spread
far and it doesn’t spread long which is
a good thing so this brings me to my
final bit I guess which is it’s just how
to manage this now all the other stuff
I’ve spoken to about so far is kind of
solid we you know we’re quite confident
in this there’s this other stuff now
we’re starting to get into we’re
starting to get some findings around how
to manage it now what we did is we
looked at threads on Twitter so if you
post a tweet and people replied to it
and they reply to it and they reply to
it and then applied to it and these can
get quite long so we looked following a
cyber hate remark how long these threads
would get and we measured again
statistically what were the factors that
meant they’d end up as long threads or
short threads the idea being can we see
how it stemmed now what we found was if
you go in there and call someone a
keyboard warrior absolutely lights a
fire the heat gets worse it gets more
extreme it gets really really very much
longer drawn-out we go in there and call
someone a racist actually they tend to
stop validation they’ve got what they
came for but what we also saw was that
if this happens in small numbers kind of
creates a bit of a hook for the person
who’s posted this original message to
come back you know someone’s kind of I
maybe validated me but someone else has
said someone else I can hook on there’s
really not much point engaging in this
at all
and the
the most useful remarks were just cause
to order you can’t do that it’s not
acceptable or requesting evidence and
there isn’t any evidence but the most
important finding is that the more
people individual people who get
involved in counter speech in this
thread trying to stop it trying to give
a message that says this is unacceptable
the shorter the thread became so
essentially what we’re seeing is
sensible online community activism so
following the riots in London right this
is an offline event but people came out
with their brushes and said look you
know we’re still here we’re going to
clean up we’re going to sort this out
you know you’re not going to beat us
this is in the offline world but the
online in the offline world can’t be
separated any more right when people are
attacked online they feel it in their
person and they feel it just in their
homes they don’t want to go out the
house they feel it when they go to work
they feel like everyone’s looking at
them just because something is online
doesn’t mean it doesn’t affect them in
the offline world we live in an online
in an offline world social cohesion is a
very big problem at the moment and cyber
hate online is there and is available
for anyone in the world to see the web
connects us all so my final remark will
be is you’re all online online social
media so if you see it work together and
you can stop it thank you very much
[Applause]
Please follow and like us: