00:00:02
make America tweet again that's the
00:00:04
provocative title of a new research
00:00:07
paper out of won that looked at the
00:00:09
Tweet surrounding the 2016 Republican
00:00:12
primary debates here to talk about the
00:00:15
paper and its surprising findings are
00:00:18
Wharton Professor Ron Burman and
00:00:20
doctoral candidate Coleman Humphrey so
00:00:23
welcome thank you deorah so tell me why
00:00:27
did you decide to analyze Twitter versus
00:00:30
other types of communication so that's
00:00:32
actually a very interesting story so our
00:00:33
co-authors professor Robert mayor and
00:00:36
shirim maluma do a PD candidate at
00:00:39
Columbia business school they were
00:00:40
starting to work on a project using
00:00:42
Twitter data um and the focus was how
00:00:45
emotionality and the device he use
00:00:47
changes the patterns but then uh we
00:00:50
realized that uh in um elections and
00:00:53
debates what happens is that people
00:00:55
today use Twitter as the main forum for
00:00:57
pushing their opinions and debating with
00:00:59
everyone Etc so this has become like the
00:01:00
Twitter election and as a result we
00:01:03
wanted to focus very much to see what do
00:01:05
voters think given what they see on
00:01:07
Twitter but also what do voters do on
00:01:08
Twitter and how they interpret that now
00:01:12
we this was about in August at that time
00:01:15
it was clear that Trump has a big
00:01:17
advantage on Twitter and he's uh getting
00:01:19
a lot of uh voter opinion and voter
00:01:22
advantage through Twitter so combined
00:01:25
with that it was very very interesting
00:01:26
to look at Twitter as does it reflect
00:01:28
voter opinion but also can you use
00:01:31
Twitter to also influence voter opinion
00:01:33
and how that works during a debate and
00:01:36
can you give us an overview a quick
00:01:38
overview of how you conducted the
00:01:40
research well sure so first um to decide
00:01:43
what data to collect we decided to focus
00:01:45
on three very pivotal debates so the
00:01:47
first one was the August debate which
00:01:50
was the most viewed primary debate of
00:01:52
all time um and people were clamoring to
00:01:56
see Trump versus bush and who would do
00:01:58
better we also um had the February
00:02:01
debate um which was just before super
00:02:04
Tuesday so it was kind of like the last
00:02:06
chance for uh one of the other
00:02:09
candidates to try and derail Trump and
00:02:10
his uh amazing poll numbers and then
00:02:13
finally the March debate which um people
00:02:16
were buzzing about because it was back U
00:02:18
Trump versus Megan Kelly the the second
00:02:20
time I guess so within there we
00:02:22
collected all the relevant uh tweets
00:02:25
from all debates so from an hour
00:02:27
beforehand to uh two hours after the
00:02:28
debates just all tweets that that had
00:02:31
the main hashtag in them so we knew they
00:02:33
were tweeting about the debate uh and
00:02:35
from there from this data we were able
00:02:36
to you know we have um who tweeted and
00:02:39
who retweeted what time that all
00:02:41
happened at and we also then we applied
00:02:44
this program called Luke which um it
00:02:47
finds was there a negative tone in the
00:02:49
Tweet was there a positive tone did you
00:02:50
use like um oops sorry did you use
00:02:52
personal pronouns um was a kind of uh
00:02:55
based on like was it a power tweet or a
00:02:57
reward tweet and then for a subset of
00:02:59
the data the kind of the most popular
00:03:01
tweets we also sent them off to MK to
00:03:05
get humans to look a little bit further
00:03:07
at the tweets we got um humans to decide
00:03:10
is there a picture or a video in the
00:03:12
Tweet or does it contain humor or
00:03:14
sarcasm uh as well so what are the key
00:03:17
findings of your paper um so so we found
00:03:21
a a few very interesting things some of
00:03:23
them just didn't conform with past
00:03:25
research results and some of them were
00:03:26
just brand new the first thing we looked
00:03:28
at was um
00:03:30
what do people see during the debate do
00:03:32
they mostly see new tweets or do they
00:03:34
see more and more old tweets because of
00:03:36
retweets of other people and we found
00:03:38
that over time you basically have a more
00:03:39
dated view of of the debate which means
00:03:42
if you look very towards the end of the
00:03:44
debate actually you're going to see
00:03:45
mostly tweets from very early on in the
00:03:46
debate the second thing we looked at was
00:03:49
what was the sentiment during the debate
00:03:50
uh for different candidates and we found
00:03:53
that if you looked after the debate
00:03:54
versus during the debate you would get a
00:03:56
very different view of the sentiment for
00:03:58
so example for Donald Trump the uh
00:04:01
sentiment after the debate was uh very
00:04:03
positive while during the debate
00:04:05
sometimes there were controversies maybe
00:04:07
Sensations that generated a negative um
00:04:11
sentiment uh and finally another thing
00:04:13
we noticed is that the Tweet stream
00:04:16
becomes more and more Sensational and
00:04:18
less substantive and in the words of of
00:04:21
Bob Mayer another professor o on this
00:04:23
colleagu is that during the debate you
00:04:25
get the New York Times after the debate
00:04:27
you getting the New York Post and this
00:04:29
is something thing we found out which
00:04:30
was very interesting do you think this
00:04:32
explains why Trump defied expectations
00:04:35
to get the Republican nomination for
00:04:38
president sure this we feel like this
00:04:41
what Rono said definitely speaks to um
00:04:43
to one of the reasons why perhaps he was
00:04:45
so popular um Twitter kind of like a lot
00:04:48
of other media I guess just really
00:04:50
focuses on tabloid or Sensational stuff
00:04:52
especially afterwards like that's that's
00:04:54
what sticks um one of the things we
00:04:57
analyzed was was sentiment people
00:04:59
positive or negative towards certain
00:05:01
candidates and you can see for Trump
00:05:03
that even as something like the Trump
00:05:05
University is being discussed um like
00:05:08
quite heavily and he's really getting
00:05:09
hammered for it in the debate his
00:05:10
sentiment just like doesn't really seem
00:05:12
to drop like uh and and especially after
00:05:14
debates you can see his sentiment
00:05:15
actually goes right back up and and is
00:05:17
generally positive after the debate so
00:05:18
people just really don't care about this
00:05:21
this the negative stuff he says just
00:05:23
doesn't seem to matter um and again
00:05:27
building on on what Ron's saying it
00:05:28
seems like Trump is the perfect cidate
00:05:30
for the the Twitter election um since he
00:05:33
is so used to generating like
00:05:34
Sensational um news and and generating
00:05:37
controversy and kind of leading towards
00:05:39
all the Tabloid stuff and and also
00:05:40
avoiding policy where he's probably
00:05:42
weakest out of all the candidates on the
00:05:44
stage and that's what sticks so
00:05:46
absolutely um Twitter seems to seems to
00:05:49
be the medium the medium for him so what
00:05:52
are the implications of your findings on
00:05:54
voter
00:05:56
opinions so there are a few implications
00:05:58
we think are are playing like are play
00:06:01
there there there implications on trying
00:06:03
to to gauge or to understand voter
00:06:05
opinions and and one of the findings is
00:06:07
that if you analyze Twitter data you
00:06:09
might get the wrong idea so if you look
00:06:11
during the debate maybe you'll get one
00:06:13
sentiment after the debate you might get
00:06:15
a different sentiment so it really
00:06:16
matters when you look at the debate and
00:06:19
the other thing is um that voters who
00:06:23
there are some voters who are very
00:06:24
activing the debate and watch it and
00:06:26
tweet about it Etc but most of the
00:06:27
people probably open the newspaper the
00:06:29
next morning or open their Twitter and
00:06:30
say what happened during the debate
00:06:32
yesterday and then they look and they
00:06:34
see this backwards view which means the
00:06:37
influence on them is probably much
00:06:40
stronger U than what you would get if
00:06:41
you watch during the debate one for
00:06:44
example one thing we notice is that
00:06:46
tweets that came from news sources
00:06:47
during the debates were very influential
00:06:49
during the debate and were retweeted a
00:06:51
lot Etc but after the debate they were
00:06:53
kind of going down if you think about it
00:06:55
it used to be that there was a debate
00:06:57
and news sources maybe would uh um kind
00:07:00
of summarize or comment on the results
00:07:03
after that and you would get what
00:07:04
happened Deb during the debate in in a
00:07:06
newspaper after that if you look at
00:07:08
Twitter you would get a very different
00:07:10
idea which means that perhaps Twitter is
00:07:13
not either a good measurement or maybe
00:07:15
Twitter is a reflecting uh Forum of the
00:07:18
results and the news media is not a good
00:07:20
measurement and reflecting uh Forum of
00:07:23
the results which might cause surprises
00:07:25
later
00:07:26
on did any conclusions surprise you so
00:07:29
one interesting uh finding we had which
00:07:31
was surprising was how media was almost
00:07:35
um inactive during the debate in
00:07:37
actually commenting on it what they were
00:07:39
doing is picking up mostly quotes from
00:07:42
TV and just tweeting them they weren't
00:07:44
saying who they think is winning they
00:07:47
weren't commenting on facts they were
00:07:50
just quoting and as a result during the
00:07:52
debate people were retweeting them a lot
00:07:54
but after the debate they had very very
00:07:56
very little influence this was one
00:07:58
surprise finding the second one is that
00:08:01
for many tweets so this comes to
00:08:03
academic research typically we assume
00:08:05
there's something called a diffusion
00:08:07
pattern you would have an idea that goes
00:08:09
up gets retweeted and goes down and dies
00:08:11
for many things we saw they barely had
00:08:14
any influence during the debate they
00:08:16
basically died during the debate and
00:08:17
then had the Resurgence after the debate
00:08:20
so for example um this um basically
00:08:23
fight almost between Megan Kelly and
00:08:26
Trump in August at the beginning
00:08:28
generated some attention during the
00:08:30
debate but then died off and but then
00:08:32
after the debate it became the primary
00:08:33
thing everyone was talking about and
00:08:36
just building on that what was also
00:08:37
perhaps surprising is that strong
00:08:40
positive or negative emotions didn't
00:08:42
didn't help much at all in fact what
00:08:43
actually in in our case hurt the
00:08:46
popularity of tweets both before and
00:08:47
after the debate um and just kind of
00:08:50
building on what we've been saying like
00:08:51
being being everything being so
00:08:52
Sensational and not being the most
00:08:54
popular stuff um uh like it's it was
00:08:58
especially clear afterwards so for
00:09:00
example beforehand one thing that
00:09:02
actually did stir some interest was the
00:09:04
economy when people some people did
00:09:05
tweet about the economy but afterwards
00:09:06
it was just nothing there was just no no
00:09:09
substantial policy not even policy that
00:09:11
was popular during the debate um this
00:09:14
was so it was a bit almost surprisingly
00:09:17
sad that
00:09:19
Twitter you might expect that Twitter is
00:09:21
finally the way for everyone to debate
00:09:23
and argue and and maybe get an
00:09:25
understanding of of the issues but
00:09:28
actually everyone towards the end was
00:09:29
focused more on Sensations which means
00:09:31
maybe Twitter is not the place you want
00:09:32
to look at and kind of understand what
00:09:35
is important in this election at at
00:09:36
least in terms of the issues so how is
00:09:39
your research different from prior work
00:09:41
in this area um so so research is
00:09:44
different on both methodology and a bit
00:09:47
on findings so methodology it's
00:09:49
interesting because we looked both at
00:09:51
who tweets and who retweets and who gets
00:09:54
gets followed by whom but also we looked
00:09:57
at the sentiment and the topic of the
00:09:58
tweets so this is one methodology
00:10:01
difference and another one is the
00:10:03
setting most people that analyzed uh
00:10:05
tweets were looking at you know are
00:10:08
people tweeting more about uh pop
00:10:10
culture versus music or something like
00:10:12
that if it's an event which is very
00:10:15
active and you get more than 200 tweets
00:10:17
a m a second right a second then there's
00:10:21
this fight for attention and everyone is
00:10:22
trying to say something this creates a
00:10:25
very different Dynamic um and we think
00:10:27
this is very unique because no one has
00:10:29
has previously analyzed Tweets in a very
00:10:31
short period of time in a very um kind
00:10:34
of energetic active uh event so how will
00:10:37
you follow up this research so building
00:10:40
on what Ron just said we we or others
00:10:41
might hope to um analyze other similar
00:10:44
situations such as um sports or um any
00:10:49
any other situation where it's very
00:10:50
highly focused not over a long period of
00:10:52
time um another area of focus we're
00:10:55
interested in is um this the phenomenon
00:10:58
of a lot of tweets being picked up by
00:11:00
this by different users and what we mean
00:11:01
by that is not just that something is
00:11:03
retweeted but is is tweeted like not a
00:11:05
retweet an original tweet um as opposed
00:11:08
to just being like a very popular tweet
00:11:10
like someone says something funny and
00:11:11
everyone thinks it's hilarious and
00:11:13
retweets so one example is is um quotes
00:11:16
So sometimes a quote is so
00:11:19
instantaneously interesting that
00:11:20
everyone picks it up and just to give
00:11:22
you one example from the debate U Mike
00:11:24
hucke in one of the debates said um uh
00:11:26
the purpose of the military is to kill
00:11:28
people and break things and and this was
00:11:30
retweeted this was tweeted excuse me not
00:11:32
retweeted tweeted out by just over a 100
00:11:34
people just immediately people were
00:11:36
thought this was fascinating it actually
00:11:37
wasn't that popular of of an idea like
00:11:40
it didn't um last for ages and ages
00:11:42
afterwards but it was initially very
00:11:43
popular as opposed to this some quotes
00:11:45
are only interesting because they get
00:11:46
brought up by a person in context so
00:11:50
actually in the same example um Pat and
00:11:52
oswal um tweeted out um kill people and
00:11:55
break things Mike Cooke just described
00:11:58
the GOP and that itself is a very
00:12:00
popular tweet and there are definitely
00:12:01
some quotes that get picked up in that
00:12:02
way just someone someone gives them more
00:12:04
context that becomes super popular or
00:12:06
sometimes just the quote itself just
00:12:07
goes on and that's not the only thing we
00:12:09
have um you know token tweets like who
00:12:11
won is of course very popular but but
00:12:13
it's it's interesting to see that in
00:12:14
certain circumstances some tweets are
00:12:17
either completely ignored of course some
00:12:19
are picked up because one one user
00:12:21
tweets something funny and that gets
00:12:23
retweet a lot and sometimes just one
00:12:25
idea or one moment who just gets picked
00:12:28
up so much like I guess another moment
00:12:29
that was a big deal was um uh the debate
00:12:32
between Rubio and Trump's uh who had the
00:12:34
bigger uh hands so to
00:12:36
speak so these are and the other thing
00:12:39
we didn't analyze so this was um a
00:12:42
debate where people were trying to VI
00:12:44
for the Republican nominations so these
00:12:46
are not opposing ideas or issues this is
00:12:49
just who gets the most attention we also
00:12:51
have data from um the the primary
00:12:54
debates right between Hillary Clinton
00:12:56
and Donald Trump and I think it would be
00:12:58
very interesting to understand if you
00:12:59
had the same phenomenons and the same
00:13:02
patterns there and would it be would
00:13:05
have have been able to allow you to
00:13:07
predict maybe who's going to win or or
00:13:11
do issues matter or Sensations matter
00:13:13
more and basically applying the same
00:13:15
findings from this research to another
00:13:18
data set to see if you get the same
00:13:20
things does it generalize or is it very
00:13:22
very specific well we can't wait to find
00:13:24
out what you discover in your new
00:13:26
research but thank you very much for
00:13:28
joining us today thank you very much d
00:13:32
[Music]
