1
00:00:00,110 --> 00:00:14,519
*34C3 preroll music*

2
00:00:14,519 --> 00:00:15,969
Herald: The next speaker is born and

3
00:00:15,969 --> 00:00:21,650
raised in Germany. He lives and works as a
PhD student in Canada as a member of a

4
00:00:21,650 --> 00:00:27,320
research group on extremist politics in
democratic systems and he'll give us an

5
00:00:27,320 --> 00:00:32,640
insight into the public discourse in
Germany focused on the so called

6
00:00:32,640 --> 00:00:37,750
"Alternative für Deutschland". Please
welcome Alexander Beyer.

7
00:00:37,750 --> 00:00:45,250
*applause*

8
00:00:45,250 --> 00:00:53,120
Beyer: Thank you very much. Thank you
people, for showing up in the Saal Borg,

9
00:00:53,120 --> 00:00:58,980
thank you, the internet, for watching, a
very big thank you for the organizers, for

10
00:00:58,980 --> 00:01:03,999
giving me the opportunity to, to give this
little talk. Yeah, my name is Alexander

11
00:01:03,999 --> 00:01:11,020
Beyer and everywhere I went this winter, I
didn't have to wear a winter jacket

12
00:01:11,020 --> 00:01:15,510
because the temperatures were very mild
and I will tell you in a minute why that

13
00:01:15,510 --> 00:01:19,600
matters.
As I already said, I'm a member of a

14
00:01:19,600 --> 00:01:27,790
research group in Vancouver, where we look
at what happens, how extremist parties and

15
00:01:27,790 --> 00:01:33,940
politics fare in democratic systems and
we decided to focus this research project

16
00:01:33,940 --> 00:01:40,229
on the fascinating - for researchers
fascinating - case of Germany and asking

17
00:01:40,229 --> 00:01:46,640
the questions, if we can point fingers and
is it a viable, a valid

18
00:01:46,640 --> 00:01:53,010
judgment to say, the media, the
media is to blame for the rise of the AfD?

19
00:01:53,010 --> 00:01:58,790
For anyone who, who decided at the end of
2017 that they would spend most of 2016 in

20
00:01:58,790 --> 00:02:02,590
hibernation - which seemed like a pretty
good idea at the time - I will give a

21
00:02:02,590 --> 00:02:09,770
quick rundown what happened: So, we had
election in September and the domino

22
00:02:09,770 --> 00:02:16,520
piece that was Germany fell. Domino piece
in a sense, that all around in Europe far-

23
00:02:16,520 --> 00:02:22,660
right parties had considerable success in
the past, in the recent past, and Germany

24
00:02:22,660 --> 00:02:26,900
was the sort of last stalwart in
Central Europe, where a far-right party did

25
00:02:26,900 --> 00:02:35,760
not get at the government, this
happened in September and it did not only

26
00:02:35,760 --> 00:02:40,750
get into government, er, into Parliament, it
also, the way that it looks like now, it

27
00:02:40,750 --> 00:02:43,760
might become the official leader of the
opposition.

28
00:02:43,760 --> 00:02:49,920
So, when these results came in, pundits
were really, really quick to call the

29
00:02:49,920 --> 00:02:55,479
shots.
The, the dominating sentiment was, that it

30
00:02:55,479 --> 00:03:00,930
was the media's fault: They took the
positions of the AfD and gave

31
00:03:00,930 --> 00:03:08,470
disproportionate amounts of coverage to
this far-right extremist party. And this

32
00:03:08,470 --> 00:03:16,350
sentiment had a lot of truthiness to it.
So, it had a lot of: "Yeah, sure, I can see

33
00:03:16,350 --> 00:03:21,380
why", right, everyone that opened his
newspaper or opened a news website,

34
00:03:21,380 --> 00:03:24,760
stories about the AfD seemed, or about
anything.. about something that's related

35
00:03:24,760 --> 00:03:31,839
to the AfD, seemed to dominate coverage.
This went along with a little bit of a

36
00:03:31,839 --> 00:03:39,760
felt truth, a truth that was perceived by
people, about how the campaigning season

37
00:03:39,760 --> 00:03:43,890
was a lot of season and not a lot of
campaigning, despite Martin Schultz's best

38
00:03:43,890 --> 00:03:48,160
efforts.
A whole lot of sunshine, but not a lot of

39
00:03:48,160 --> 00:03:55,140
conflict and this was something that then
was perceived to be very, uh, well, I don't

40
00:03:55,140 --> 00:04:04,150
want to say very skillfully, but somehow
filled by the AfD and and the topics that

41
00:04:04,150 --> 00:04:10,520
are of concern to this party.
So what are we doing today here? First

42
00:04:10,520 --> 00:04:17,640
off, I'm a political scientist by trade
and political scientists like theory. I

43
00:04:17,640 --> 00:04:22,431
know that this is an event, where [I] figure
you might not at the forefront of

44
00:04:22,431 --> 00:04:27,499
everyone's minds, but it is for me,
because talking and arguing to political

45
00:04:27,499 --> 00:04:32,550
scientists about theory is kind of like
mud-wrestling with a pig: you do that for

46
00:04:32,550 --> 00:04:39,270
two or three hours and then you realize,
oh, this pig actually enjoys this. So I'll

47
00:04:39,270 --> 00:04:45,960
be sort of, I have one slide on what we,
what previous theories would suggest

48
00:04:45,960 --> 00:04:50,689
has, have happened and how it could have
happened and then I'll show you what kind

49
00:04:50,689 --> 00:04:56,520
of data we have collected, to
systematically answer this question and

50
00:04:56,520 --> 00:05:02,039
talk about public discourse in Germany and
then to the meat and potatoes of the talk,

51
00:05:02,039 --> 00:05:09,090
about how the campaign unfolded in the
media and I will than, to end I will show

52
00:05:09,090 --> 00:05:15,169
some more data, that is a bit different,
that paints a picture on why this election

53
00:05:15,169 --> 00:05:21,800
was a special election and why it was sort
of a perfect storm of an election for a

54
00:05:21,800 --> 00:05:29,919
far-right party and why this actually
makes us claim that the media could be

55
00:05:29,919 --> 00:05:35,529
said to have behaved pretty reasonable. As
a little teaser.

56
00:05:35,529 --> 00:05:44,259
OK. Theory. One slide. Two possible
mechanisms of media effects. There's this

57
00:05:44,259 --> 00:05:52,189
normative, very endearing and wonderful
idea, that if you read something, that

58
00:05:52,189 --> 00:05:59,849
someone carefully crafts and he or she
constructs an argument, that is well

59
00:05:59,849 --> 00:06:05,999
written, well made, you read this, you take
it in, you're persuaded by that,

60
00:06:05,999 --> 00:06:13,409
regardless of what this argument is. 60
years of media research suggests, that this

61
00:06:13,409 --> 00:06:17,330
doesn't happen.
Pre-existing opinions are extremely

62
00:06:17,330 --> 00:06:21,930
difficult to change in each and every
single one of us, even though we're likely

63
00:06:21,930 --> 00:06:28,089
to admit that "No, no sure, I'm a rational
thinker, I take standpoints if they're

64
00:06:28,089 --> 00:06:35,770
convincing to me and I internalize them.",
but it's not how this works.

65
00:06:35,770 --> 00:06:42,490
The second possible effect, and the one
that will be of of concern to us today at

66
00:06:42,490 --> 00:06:50,560
the core of the presentation, is something
that's called Priming. So, the media can't

67
00:06:50,560 --> 00:06:55,009
tell people what to think, it can't
persuade people independently of the

68
00:06:55,009 --> 00:07:00,639
previous opinions that people have, but
it's really, really successful in telling

69
00:07:00,639 --> 00:07:05,860
people, what to think about. It's super
good, the media is super good, reading

70
00:07:05,860 --> 00:07:11,500
something is very effective in bringing
something to the front of your mind.

71
00:07:11,500 --> 00:07:18,269
And here I.. here I can tell you, why I
told you about my my choice of attire in

72
00:07:18,269 --> 00:07:22,059
winter. A vast majority of you probably
thought when I said this "'Oh, I didn't

73
00:07:22,059 --> 00:07:27,909
have to wear a winter jacket', wow, what's..
who is this guy?" But maybe a few of you

74
00:07:27,909 --> 00:07:31,510
thought "Yeah, sure it was pretty mild,
that's climate change."

75
00:07:31,510 --> 00:07:37,969
So, without naming the issue, I..
there's a chance, that I primed a few of

76
00:07:37,969 --> 00:07:43,169
you, to consider climate change and pull
that in your frontal lobe, at the front of

77
00:07:43,169 --> 00:07:49,490
your mind.
This is, this is important, this is a.. the

78
00:07:49,490 --> 00:07:57,050
central thing, that we have to consider if
we ask, if the media wrote up a party like

79
00:07:57,050 --> 00:08:03,259
the AfD.
Also important to consider here is, that

80
00:08:03,259 --> 00:08:10,780
priming is easier - or there's an indirect
effect of priming as well - where a topic

81
00:08:10,780 --> 00:08:17,029
that is owned by a specific party, that's
the thing that then favors the party

82
00:08:17,029 --> 00:08:20,749
subsequently.
So, if the media writes a lot about

83
00:08:20,749 --> 00:08:28,589
refugees, a xenophobic far-right party,
that has this problem of refugees at the

84
00:08:28,589 --> 00:08:38,360
core of their agenda, will reap in benefits
in our minds, in that it's agenda will be,

85
00:08:38,360 --> 00:08:44,911
will fall on fertile ground.
So far the theory, that's all. So, what did

86
00:08:44,911 --> 00:08:51,040
we do based on this theory? We collected
data, lots of data, we have.. we understand this

87
00:08:51,040 --> 00:08:56,740
data.. we understand this text, that we
collected to be data and we use natural

88
00:08:56,740 --> 00:09:00,670
language processing to analyze that.
Natural Language Processing basically

89
00:09:00,670 --> 00:09:04,900
means, that we're giving language to a
computer that wasn't written specifically

90
00:09:04,900 --> 00:09:12,100
to be understood by a computer and try to
extract meaningful analysis based on what

91
00:09:12,100 --> 00:09:17,829
the computer is doing with this.
So, we used some sifting methods to

92
00:09:17,829 --> 00:09:25,770
collect about 8.500 articles from four
central German news websites: Focus, Bild,

93
00:09:25,770 --> 00:09:31,860
Welt and Spiegel. And we have.. that
results in a unique data set, that, to our

94
00:09:31,860 --> 00:09:39,250
knowledge, no one else has. If so, please
reach out to us.

95
00:09:39,250 --> 00:09:45,940
And this was so unique, that it deserves at
least six fire emojis. It was also pretty

96
00:09:45,940 --> 00:09:51,350
exciting because, that was pretty cheap. We
were two people that were mainly concerned

97
00:09:51,350 --> 00:09:56,639
with collecting this data and I don't want
to, I don't want to calculate my hourly

98
00:09:56,639 --> 00:10:03,750
wage, but it was almost done with no
financial expense. And this is cool,

99
00:10:03,750 --> 00:10:08,740
because we're social scientists, we're
faced with this problem, with this very

100
00:10:08,740 --> 00:10:15,899
interesting case of Germany sort of
falling in line, very delayed, with lots of

101
00:10:15,899 --> 00:10:20,430
other countries around it - in terms of
the far-right party getting their seats in

102
00:10:20,430 --> 00:10:25,230
Parliament and we can use methods that are
available to us, if we're sort of like

103
00:10:25,230 --> 00:10:27,620
sitting down and reading our stack
overflow and sort of teaching those

104
00:10:27,620 --> 00:10:32,039
methods to us, to systematically try to
answer this question.

105
00:10:32,039 --> 00:10:38,680
Let's dive right in. The share of party
mentions in online news. So, what we did

106
00:10:38,680 --> 00:10:45,980
for each day, we calculated what the total
number of mentioned political actors is.

107
00:10:45,980 --> 00:10:51,990
We did that based on word lists, that we
carefully crafted, that included candidate's

108
00:10:51,990 --> 00:10:56,770
names and party abbreviations and party
names and things like "Kanzlerin" and

109
00:10:56,770 --> 00:11:05,670
"Kanzlerkandiat" for the CDU/CSU and the
SPD respectively and we let that thing rip

110
00:11:05,670 --> 00:11:12,700
through our little rscript that we have.
So, the average of mentions of each party

111
00:11:12,700 --> 00:11:19,269
over the course of the campaign looks
something like this: Between July 1st and

112
00:11:19,269 --> 00:11:25,930
September 24th - that's the time frame
that we concentrated on - we see a clear

113
00:11:25,930 --> 00:11:31,790
incumbency bonus, the "Kanzlerbonus" the
"Kanzlerinbonus" for the CDU/CSU, social

114
00:11:31,790 --> 00:11:40,100
democrats high twenties, and the AfD at
10.7 %. Here we might say at smaller

115
00:11:40,100 --> 00:11:44,450
parties.. a little note to the green and
the left, so with this dictionary method is

116
00:11:44,450 --> 00:11:48,990
kind of tricky, because we can't say: "Oh yeah,
well we just gonna count every occurrence

117
00:11:48,990 --> 00:11:55,379
of Grüne and Linke" for the Green Party and
the Left Party, because then we get stuff

118
00:11:55,379 --> 00:12:00,529
like 'the green banana' and 'the left hand'
that is counted for them. So that's why

119
00:12:00,529 --> 00:12:03,960
here we're only using candidates' names.
That's why they probably.. they sort of

120
00:12:03,960 --> 00:12:09,329
underperform. But for our purpose of
talking about why the AFD got favored by

121
00:12:09,329 --> 00:12:14,360
the media, we're sort of letting that drop
under the table. So, the story here is

122
00:12:14,360 --> 00:12:21,740
over the course of the campaign, 10.7 % of
mentions were happening that mentioned the

123
00:12:21,740 --> 00:12:28,750
AFD, basically. Case closed. Right, AFD got
12.7 % in the election. That doesn't really

124
00:12:28,750 --> 00:12:36,040
sound like it was favored by the media.
And a few of you might know this analysis

125
00:12:36,040 --> 00:12:42,509
from a blog post, that me and Constanze
Kurz wrote for Netzpolitik, sort of like 45

126
00:12:42,509 --> 00:12:47,080
seconds after the election, when we worked
on truncated data. And we also focused on

127
00:12:47,080 --> 00:12:54,720
print media and this is sort of what this
graph looked like, that we based our

128
00:12:54,720 --> 00:12:59,480
conclusion on. AFD didn't really get any
disproportionate amount of coverage. It

129
00:12:59,480 --> 00:13:02,480
actually is in the.. in the last week of the
campaign.. last weeks of the campaign

130
00:13:02,480 --> 00:13:10,410
actually is outperformed by the FDP.
Science is the current state of airing, or

131
00:13:10,410 --> 00:13:18,370
the.. so, now that we have better data in
terms of online news data this whole story

132
00:13:18,370 --> 00:13:23,090
looks a bit different. If we take the
average over the whole course of the

133
00:13:23,090 --> 00:13:29,139
campaign and actually have it shown to
us.. Stay by the day - This is what I want

134
00:13:29,139 --> 00:13:36,029
to focus on now.
So just looking at the sort of tail end of

135
00:13:36,029 --> 00:13:41,089
this all the way to the right, when we get
close to the election date, the order of

136
00:13:41,089 --> 00:13:46,749
this is surprisingly close to the actual
election results. The parties actually do

137
00:13:46,749 --> 00:13:58,379
get in the order, that they came out of
the election. But we do see a little curve

138
00:13:58,379 --> 00:14:04,899
that gets closer to a curve that should be
bigger. And this is where the.. well, I

139
00:14:04,899 --> 00:14:09,829
don't want to say magic, but this is
where the interesting stuff lies. So let's

140
00:14:09,829 --> 00:14:16,639
look at the curves one after the other.
The CDU/CSU, as you would expect, as the

141
00:14:16,639 --> 00:14:21,589
incumbent anything that is remotely
political in domestic and international

142
00:14:21,589 --> 00:14:26,569
politics, will score mentions for the
Chancellor and the CDU/CSU. That's why

143
00:14:26,569 --> 00:14:30,480
this curve is considerably higher than the
others, but we do see a downward tendency

144
00:14:30,480 --> 00:14:34,940
the closer we get to the campaign, when
campaign coverage shifted from the

145
00:14:34,940 --> 00:14:41,220
incumbent to the competitors. Especially
the underdog competitors. Which is kind..

146
00:14:41,220 --> 00:14:47,670
that's a bad transfer to the SPD
now, but if we look at the curve of the

147
00:14:47,670 --> 00:14:52,831
Social Democratic Party, there's a slight
bump around August and Martin Schulz

148
00:14:52,831 --> 00:14:59,749
really tried to drive home this issue of
justice as the central campaign promise

149
00:14:59,749 --> 00:15:02,870
and there's another little slight
humper on September 1st, begin of

150
00:15:02,870 --> 00:15:08,430
September, when the televised debate
happened, but the overall trend is pretty

151
00:15:08,430 --> 00:15:13,459
linear, doesn't seem to be, if we would
just smooth this plot out to be a straight

152
00:15:13,459 --> 00:15:20,430
line, it probably would be pretty much
horizontal. Not so for the AFD. So

153
00:15:20,430 --> 00:15:25,110
remember, over the course of the campaign
they got 10.7 % on average of mentions

154
00:15:25,110 --> 00:15:28,930
*drinks from his bottle*
And that's true. If we calculate an

155
00:15:28,930 --> 00:15:35,530
average of that, of course, this looks like
it scores considerably lower than the two

156
00:15:35,530 --> 00:15:42,100
major parties. But something happens in
late August and all of a sudden this party

157
00:15:42,100 --> 00:15:46,699
gets actually close to the Social
Democrats. It like.. starting in late

158
00:15:46,699 --> 00:15:52,689
August, the tendency becomes one, that is
pretty considerably upwards. And if we

159
00:15:52,689 --> 00:15:57,670
take the average of only the two last
weeks before the election, we get to a

160
00:15:57,670 --> 00:16:05,029
number of 19.6 of all mentions are talking
about the AFD there. Which is something, if

161
00:16:05,029 --> 00:16:11,019
you think about the mechanisms of priming,
those are short term effects. We're

162
00:16:11,019 --> 00:16:14,389
looking for things that happen over a
short term or have an effect in a pretty

163
00:16:14,389 --> 00:16:20,570
short term. So this is something that is
extremely important. At the beginning of

164
00:16:20,570 --> 00:16:24,650
this time frame, where the plot becomes
something that has a trend that shows

165
00:16:24,650 --> 00:16:30,669
upwards, around, like, August 28th - where
that first little mountain.. first little

166
00:16:30,669 --> 00:16:35,930
summit occurs - two things happened: one, a
refugee boat capsized in the

167
00:16:35,930 --> 00:16:41,170
Mediterranean. An event that we've sadly
and.. have to see terrifyingly often and

168
00:16:41,170 --> 00:16:45,199
one of the people died. And the second
thing that happened was, that Alexander

169
00:16:45,199 --> 00:16:50,990
Gauland in an interview claimed, that a
German politician should be dumped in

170
00:16:50,990 --> 00:17:02,550
Anatolia. And it's interesting, if you talk
about.. if you extract the topics, that

171
00:17:02,550 --> 00:17:07,508
are covered in relation to the AFD before
and after this moment.

172
00:17:07,508 --> 00:17:15,329
Before this August 28th, it's a lot about
Alice Weidel writing emails where it turns

173
00:17:15,329 --> 00:17:20,380
out, she's not the public persona that she
claims she is and it's a lot about

174
00:17:20,380 --> 00:17:26,409
internal rifts of this far-right party.
The internal tensions between the super

175
00:17:26,409 --> 00:17:35,029
far right wing and the far right or right
wing-wing and afterwards, there's a

176
00:17:35,029 --> 00:17:41,610
surprising amount of citations of this "Oh
we're gonna... we should dump this person

177
00:17:41,610 --> 00:17:47,179
in another country." So that's something
that indicates, that this strategy of sort

178
00:17:47,179 --> 00:17:56,360
of provoking a scandal paid off. But
let's.. before we get into that, let's

179
00:17:56,360 --> 00:17:59,740
look into the topics that were covered
over the course of the campaign. We did

180
00:17:59,740 --> 00:18:05,270
the same thing, we developed topic
dictionaries with keywords for each

181
00:18:05,270 --> 00:18:09,530
category and we let our script read
through all the data and count

182
00:18:09,530 --> 00:18:18,520
occurrences. So looking at this, we see a
sort of band there in the 10% range,

183
00:18:18,520 --> 00:18:24,640
where it's all a colourful rainbow, where
the topics don't really diverge from each

184
00:18:24,640 --> 00:18:29,241
other, except for that
topic of domestic security, which

185
00:18:29,241 --> 00:18:34,150
is there at the low end of the range. But
we do have one topic, that stands out quite

186
00:18:34,150 --> 00:18:39,700
considerably in the early months of the
Wahlkampfsommer, which is European Union,

187
00:18:39,700 --> 00:18:47,010
generally European Union topic. This is
because on July 1st Helmut Kohl, the

188
00:18:47,010 --> 00:18:51,950
eternal Chancellor, get the first European
act of state and a lot of things were

189
00:18:51,950 --> 00:18:55,909
written about his legacy in terms of the
European Union and lots of people showed

190
00:18:55,909 --> 00:19:00,249
up from Strasbourg and Brussels and paid
their respects. This is why this topic

191
00:19:00,249 --> 00:19:05,990
seems like.. or this is, why this topic comes
in as strong as it does here. Now the

192
00:19:05,990 --> 00:19:10,590
topic that has a sort of unusual curve
here on our graph, is the topic of the

193
00:19:10,590 --> 00:19:17,679
environment. Our dictionaries that we
developed were topical and so what causes

194
00:19:17,679 --> 00:19:26,090
this steep, steep summit there in early
August, is the Dieselgipfel.. there, the

195
00:19:26,090 --> 00:19:31,700
Diesel Summit, where German car
manufacturers try to sort of get out of

196
00:19:31,700 --> 00:19:38,669
the fact, that they basically ripped off
customers with selling cars that emitted

197
00:19:38,669 --> 00:19:44,250
toxic amounts of poisonous gas and dust.
This is why this is extremely important in

198
00:19:44,250 --> 00:19:49,000
the high.. in the low 40 % range in
early August. But afterwards the trend

199
00:19:49,000 --> 00:19:55,370
line points steeply down.
A topic that was pretty consistent over

200
00:19:55,370 --> 00:20:01,580
the course of the campaign in its overall
dynamic or at the sort of.. not the

201
00:20:01,580 --> 00:20:06,460
overall dynamic, but the role that it
played, is the topic of immigration. And

202
00:20:06,460 --> 00:20:13,909
immigration means migration and refugees
in our case here. And now, thinking about

203
00:20:13,909 --> 00:20:19,980
what that means in relation to our theory
on priming, we would think that sure,

204
00:20:19,980 --> 00:20:25,580
that's a topic that is owned by the AFD.
It's, like, it's super tightly connected to

205
00:20:25,580 --> 00:20:34,360
that party's rise. So, this is something
that does favour a far-right party like

206
00:20:34,360 --> 00:20:40,890
those are, like it is. But we can do
a sort of more systematic investigation

207
00:20:40,890 --> 00:20:50,710
into this. So, this graph shows you the
poles: each dot represents polling results

208
00:20:50,710 --> 00:20:55,900
for the AFD and the line is the average
out of those polls, again over the course

209
00:20:55,900 --> 00:20:59,990
of the of the time frame that we surveyed.
Pretty much constant

210
00:20:59,990 --> 00:21:03,840
until mid-August, and all of a sudden we have
increasing variance and we have a tendency,

211
00:21:03,840 --> 00:21:11,100
a trend line, that points upwards. And now,
this is where the heart of the story lies:

212
00:21:11,100 --> 00:21:19,280
is this, is this dependent on the
mentions that the AFD got in the media?

213
00:21:19,280 --> 00:21:23,010
There's the orange line, now we have a
sort of we have a different, a different

214
00:21:23,010 --> 00:21:26,470
scale of our graph, that's why it looks
way more nervous than in the bigger one,

215
00:21:26,470 --> 00:21:34,669
that we had. Difficult to say. If you have
data like this, time serious data, you

216
00:21:34,669 --> 00:21:38,850
actually want to get rid of trends, in
terms of what the analysis should be like.

217
00:21:38,850 --> 00:21:44,490
So one way to do this in a graphic
representation, is by not showing the

218
00:21:44,490 --> 00:21:49,590
absolute values and how they develop, but
only showing the change from day to day

219
00:21:49,590 --> 00:21:57,490
and plotting that. This is what this graph
does. So here, these two lines dance around

220
00:21:57,490 --> 00:22:02,390
the zero mark, because - especially the blue
one, where it's the polling results - there

221
00:22:02,390 --> 00:22:06,141
wasn't a lot of variation from day to day.
It's in incremental steps that the curve

222
00:22:06,141 --> 00:22:13,370
points up and down. It gets a bit more, a
bit higher in variance around the.. after

223
00:22:13,370 --> 00:22:20,289
the mid of August. And whereas the AFD
mentions in the media, they stay, well, they

224
00:22:20,289 --> 00:22:26,870
stay rich in variance. Hard to tell, if
anything systematic is there. You would

225
00:22:26,870 --> 00:22:33,390
think that after the sort of first, 3rd
of August, those, those lines are

226
00:22:33,390 --> 00:22:38,890
connected. We ran an analysis - a vector
autoregression model, time series

227
00:22:38,890 --> 00:22:45,159
statistics - we couldn't find any
systematic relation in a timeframe, that

228
00:22:45,159 --> 00:22:48,790
made sense for our theory on priming.
Which is a few days that we're looking

229
00:22:48,790 --> 00:22:53,770
for. So if you talk about time series, we
talk about lag and lead, and so you try to

230
00:22:53,770 --> 00:22:58,640
connect a data point that is further down
the line with a data point that is, that is

231
00:22:58,640 --> 00:23:07,549
not as far down the line, and nothing of
statistical significance showed up here.

232
00:23:07,549 --> 00:23:10,960
And this kind of stumped us - we thought,
right, when we looked at this there was

233
00:23:10,960 --> 00:23:18,510
something. That we.. we sort of took a step
back and we considered another possibility

234
00:23:18,510 --> 00:23:29,399
to.. as why to.. as to why the media reported
as they did. Did the media just give

235
00:23:29,399 --> 00:23:37,659
the people what the people wanted?
And here is why I want to talk

236
00:23:37,659 --> 00:23:43,470
to you about why this was a
special election. This graph - I adapted

237
00:23:43,470 --> 00:23:48,190
this graph from the Berliner Morgenpost
and they based it on on surveys conducted

238
00:23:48,190 --> 00:23:54,380
by Infratest dimap on to the.. the data
I didn't have any access to them. But this

239
00:23:54,380 --> 00:24:00,260
impressively shows, why there was a special
election. In five out of the six preceding

240
00:24:00,260 --> 00:24:05,350
elections, employment was the topic that
was on top of people's minds, when they

241
00:24:05,350 --> 00:24:10,389
made the decision in terms of which party
to vote for. And employment means

242
00:24:10,389 --> 00:24:17,749
unemployment. In 2017, with unemployment
being at record lows, and after 2015

243
00:24:17,749 --> 00:24:23,720
having.. or having a Syrian civil war still
going on, we're having ...

244
00:24:23,720 --> 00:24:32,300
refugees come into.. into Western and into
Europe, immigration jumps on out as the.. as

245
00:24:32,300 --> 00:24:38,679
the topic that was the most important for
people. And here, if we if we look at also

246
00:24:38,679 --> 00:24:44,760
the topics that are further down the
important scale for voters, those are all

247
00:24:44,760 --> 00:24:50,830
topics, where one could conceivably think
that those can be spun in a way that they

248
00:24:50,830 --> 00:24:58,040
are connected to this refugee situation.
Social injustice, economic injustice, that's

249
00:24:58,040 --> 00:25:02,340
something that a party like the AFD can
very effectively turn into an idea on

250
00:25:02,340 --> 00:25:10,560
group based conflict: "It's us versus them".
Same with pensions: "Oh, those people come

251
00:25:10,560 --> 00:25:15,320
here to take our jobs and our money, and
especially from the old people. From our

252
00:25:15,320 --> 00:25:26,460
elderly. So 2017, die Bundestagswahl 2017,
is a special case, if we consider it

253
00:25:26,460 --> 00:25:33,290
compared to other parties. So now having
this situation where we find that it's,

254
00:25:33,290 --> 00:25:36,580
it's something that basically never
happened in recent history and in Germany

255
00:25:36,580 --> 00:25:42,730
before in terms of what, what made people
decide at the polls. We wondered, OK,

256
00:25:42,730 --> 00:25:48,389
well, is there a way to more accurately
measure this demand side of things, this

257
00:25:48,389 --> 00:25:55,289
this need for information for.. of voters.
And what better way there is to measure

258
00:25:55,289 --> 00:26:03,510
some.. to measure the salience in the
population than to look at Google queries?

259
00:26:03,510 --> 00:26:10,260
So we collected Google Trends data - more
specifically, the Google searches on

260
00:26:10,260 --> 00:26:13,979
refugees, 'Flüchtlinge', 
general term,

261
00:26:13,979 --> 00:26:17,780
and again, here's this..
this way to even out a trend

262
00:26:17,780 --> 00:26:25,191
line - this is the daily change in
how this topic developed. And if we put

263
00:26:25,191 --> 00:26:31,769
our daily change of AFD mentions over
that, we do see that there's something

264
00:26:31,769 --> 00:26:40,370
there. There's some sort of systematic
relationship. And then, crunching these

265
00:26:40,370 --> 00:26:44,700
numbers and putting them again through a
vector autoregression model, we come to

266
00:26:44,700 --> 00:26:52,970
the conclusion, that with a lag of only one
day, Google searches for refugees actually

267
00:26:52,970 --> 00:26:58,559
lead AFD mentions in the media. So, if on
Tuesday a higher number of people in

268
00:26:58,559 --> 00:27:06,760
Germany googled "Refugees", on Wednesday, the
AFD was mentioned more often than the day

269
00:27:06,760 --> 00:27:12,429
before. The end effect wasn't big, but it
was there and it was significant. We also, of

270
00:27:12,429 --> 00:27:17,690
course, considered the alternative, and the
magic word is here it's.. it's Granger

271
00:27:17,690 --> 00:27:24,169
causality, so you can actually calculate,
and reliably calculate, the temporal

272
00:27:24,169 --> 00:27:34,559
succession, means that one follows the
other. And so all of a sudden, it becomes

273
00:27:34,559 --> 00:27:40,809
a bit difficult to point the fingers at
the media. Because, if the media just

274
00:27:40,809 --> 00:27:46,460
reacts to an interest, it operates like a
business. If we like it or not. There's

275
00:27:46,460 --> 00:27:50,240
the normative idea of the media, especially
in a country that is rich in high-quality

276
00:27:50,240 --> 00:27:57,230
publications as is Germany, that the media
is a public good, that educates people and

277
00:27:57,230 --> 00:28:01,980
brings out the best in them, in
challenging them, and persuading them of

278
00:28:01,980 --> 00:28:05,760
the best side of the argument. But at the
end of the day, in your online worlds, it

279
00:28:05,760 --> 00:28:09,250
is a business with a measurable outcome.
You have clicks, you have trackers, we

280
00:28:09,250 --> 00:28:14,730
have ad durations that you can measure.
And so you can see, which articles are

281
00:28:14,730 --> 00:28:22,770
favored and which articles people
last the longest on. And we're not saying

282
00:28:22,770 --> 00:28:26,000
- there's important distinction to make
here - we're not saying, that there's a

283
00:28:26,000 --> 00:28:30,730
direct causal link between people googling
refugees and the media directly

284
00:28:30,730 --> 00:28:36,650
reacts to that prompt, because there's
some search engine optimization guy or

285
00:28:36,650 --> 00:28:43,309
girl.. every media publishing house,
that monitors what people are interested

286
00:28:43,309 --> 00:28:48,010
in. We're saying, that there's an
intermediate step there. It's not a

287
00:28:48,010 --> 00:28:55,029
direct cause, it's just a sort of delay,
that is in there, that allows for other

288
00:28:55,029 --> 00:29:00,850
mechanisms to get in. So we're
wondering: What about the consumer focusing

289
00:29:00,850 --> 00:29:08,029
on the demand side? And in 2017, there's a
few things that you could actually look at

290
00:29:08,029 --> 00:29:18,269
to gauge what the demand-side demands, and
we decided to focus on Twitter. Because,

291
00:29:18,269 --> 00:29:21,029
without actually knowing this, when we
first started out with collecting all this

292
00:29:21,029 --> 00:29:27,220
data, we decided to set up, yeah, to set up
a Twitter scraper. And that way, between

293
00:29:27,220 --> 00:29:32,640
September 1st and September 24th, we
collected 4.5 million tweets, that

294
00:29:32,640 --> 00:29:40,110
contained keywords.. that contain any one of
a list of keywords, that had

295
00:29:40,110 --> 00:29:50,179
politic connotation. So looking at this
body of data, we can extract things like

296
00:29:50,179 --> 00:29:58,460
the top 200 most used Hashtags. And if we
do that and we, we count the tweets that

297
00:29:58,460 --> 00:30:04,540
contains one of the top 200 Hashtags and
we pay special attention, to which one of

298
00:30:04,540 --> 00:30:12,759
these Hashtags are decidedly pro AFD, we
get to a number, that 30.9 % of the tweets

299
00:30:12,759 --> 00:30:17,549
that contained any of those top 200
Hashtags, actually contain one that is in

300
00:30:17,549 --> 00:30:24,090
favor of the AFD. Whereas if we count the
decidedly no AFD, the anti AFD, no AFD in

301
00:30:24,090 --> 00:30:30,759
all ways of spelling and capitalization
and so forth, that's only 1.2 %.

302
00:30:30,759 --> 00:30:38,779
And here it becomes a bit
ticklish. So, in order to sort of give a

303
00:30:38,779 --> 00:30:46,020
better idea, of what role Twitter might
have played in our little, in our little

304
00:30:46,020 --> 00:30:51,049
relationship here, between the demand side
and the supply side, the supply side

305
00:30:51,049 --> 00:30:59,281
supplying the news, we have a
beautiful network graph. So this is a

306
00:30:59,281 --> 00:31:08,140
retweeting network: this is we extract all
the mentions of a.. of an actor. Each dot is

307
00:31:08,140 --> 00:31:13,610
a Twitter user, each line is 5 or more
retweets. Retweets, we're aware of that,

308
00:31:13,610 --> 00:31:17,219
Retweets don't automatically mean
endorsement - you might retweet something

309
00:31:17,219 --> 00:31:24,549
that is outlandish and crazy. But for the
sake of visualizing, what the weights are

310
00:31:24,549 --> 00:31:29,590
on Twitter, we're treating them as the
same. And anyone, who ever has worked

311
00:31:29,590 --> 00:31:31,970
with network graphs of that size, that
take a long time

312
00:31:31,970 --> 00:31:36,840
to generate, and it's kind of
tough to to label them, so I'm very proud,

313
00:31:36,840 --> 00:31:43,750
that I was able to do so. If we look at
this Island down there, that blob, that

314
00:31:43,750 --> 00:31:53,640
blue blob - those are accounts, that
cluster around AFD accounts. The coloring

315
00:31:53,640 --> 00:31:58,490
here was done by a walk trap
algorithm, I just adjusted the colors that

316
00:31:58,490 --> 00:32:03,139
that algorithm used to actually match the
colors in the, in the German party

317
00:32:03,139 --> 00:32:11,000
landscape. And so we do have a hefty
continent at the bottom right, that

318
00:32:11,000 --> 00:32:16,730
connects all kinds of people to the
AFD. There, if you look at the little

319
00:32:16,730 --> 00:32:24,010
appendix below here, that is colored in
brown, that is mainly organized around a

320
00:32:24,010 --> 00:32:31,840
movement called Reconquista Europe, which
is an even further right wing, right wing

321
00:32:31,840 --> 00:32:40,840
movement that is sort of, like, directly
tacked to this island of the AFD, and the

322
00:32:40,840 --> 00:32:47,759
the connecting node is Björn Höcke, which
is quite interesting. So we have the AFD down

323
00:32:47,759 --> 00:32:52,659
there, we have the other parties up there, the
rainbow, that is the pluralistic political

324
00:32:52,659 --> 00:32:59,749
landscape, we have those.. those two extreme
points there, at the at the super top right

325
00:32:59,749 --> 00:33:08,109
and there at the bottom left - that is..
those are very extremely.. extreme Twitter

326
00:33:08,109 --> 00:33:13,590
user parties. That's the ÖDP and the
Freien Wähler, so they don't.. they don't seem to

327
00:33:13,590 --> 00:33:19,370
engage with the nodes, that are in the
center here. But what's also valiable to

328
00:33:19,370 --> 00:33:25,059
note, is that for the other parties, for the
established parties, starting from the left

329
00:33:25,059 --> 00:33:30,480
and orange - the Pirate Party and then red
the Social Democrats, Purple is Die Linke,

330
00:33:30,480 --> 00:33:37,159
green die Grünen, yellow FDP and black the
Conservative Party CDU/CSU. All of these

331
00:33:37,159 --> 00:33:44,029
parties have a central node - a central, a
central account, around which a lot of

332
00:33:44,029 --> 00:33:50,389
other users are fanned out. So there's.. for
each party, there's a smaller number, or a

333
00:33:50,389 --> 00:33:55,590
relatively small number, of accounts, that
are highly favored in how often they are

334
00:33:55,590 --> 00:34:04,259
retweeted. AFD doesn't have that. Even.. so,
this is of course a projection of

335
00:34:04,259 --> 00:34:07,309
something that's 3D in a
2D place, so there might be

336
00:34:07,309 --> 00:34:11,239
some skewing going on here, in terms of how
it shows on our screen, but even turning it

337
00:34:11,239 --> 00:34:18,790
and trying to identify which party is at
the center, wasn't, wasn't really possible.

338
00:34:18,790 --> 00:34:26,010
So the internal rifts, and the internal
power struggles - they do show in how

339
00:34:26,010 --> 00:34:32,360
members of the party are retweeted. Also
interesting to note is which nodes, which

340
00:34:32,360 --> 00:34:38,770
users are connecting these two continents,
so to speak. One is, that blue dot, is

341
00:34:38,770 --> 00:34:41,860
wahlrecht.de, a polling
aggregator, of course, everyone is

342
00:34:41,860 --> 00:34:45,840
interested in getting their polling
numbers out. And there's.. that's tough to

343
00:34:45,840 --> 00:34:53,250
see here, but there's a beige user in
the middle therem which is welt.de, so

344
00:34:53,250 --> 00:34:58,020
one of the media.. one of the media
publications, that we actually collected

345
00:34:58,020 --> 00:35:02,970
data on and surveyed. Another thing that
is.. I'm just gonna mention here briefly, is

346
00:35:02,970 --> 00:35:13,830
the.. that light pink colored insert between
the greens and the central gray beige dot

347
00:35:13,830 --> 00:35:17,510
- those are Jan Böhmermann, die Heute-show
und Extra 3.

348
00:35:17,510 --> 00:35:26,460
*laughing, clapping*
Yeah, so there's a.. the dynamics are

349
00:35:26,460 --> 00:35:32,750
clear, that we have this party that is
pretty well organized on social media,

350
00:35:32,750 --> 00:35:40,280
and thus is able to dominate a media
agenda, that is based on algorithms basically.

351
00:35:40,280 --> 00:35:46,980
If you think about how, how the logic of
information dissemination works on

352
00:35:46,980 --> 00:35:54,700
Twitter: with trending Hashtags. If you
have a party that is as - well, I don't want

353
00:35:54,700 --> 00:35:58,700
to say organized - but as tightly clustered
around itself.. within itself as the AFD

354
00:35:58,700 --> 00:36:05,430
shows up here, there's a good chance, that
that will influence, what all of us get to

355
00:36:05,430 --> 00:36:11,800
see, when we check out the Twitter homepage.
Now I know, that probably a good chunk of

356
00:36:11,800 --> 00:36:18,080
you have burning questions in their mind,
and I'm gonna going to want to know - so

357
00:36:18,080 --> 00:36:25,730
how many of these of these bright blue bar
blobs are bots? Are Twitter bots. We tried to

358
00:36:25,730 --> 00:36:31,510
find that out, using a tool called the 
Botometer, which is something that has an

359
00:36:31,510 --> 00:36:35,740
API available online, where you can submit,
it's a project from a research team

360
00:36:35,740 --> 00:36:40,510
in Indiana, where you can submit the name
of a Twitter user and then it gives you a..

361
00:36:40,510 --> 00:36:46,320
it rends lots of lots of analyses and
analysis lots of things

362
00:36:46,320 --> 00:36:51,610
about this user: the frequency of tweets,
the time at which it tweets, who is.. who

363
00:36:51,610 --> 00:36:56,170
is it following, who is it followed by, who
is it talking to, that kind of stuff. But

364
00:36:56,170 --> 00:37:03,640
when I tried to submit that, I broke their
API. And so, if they're happen to watch I

365
00:37:03,640 --> 00:37:09,900
apologize, that was me. So it wasn't be
able to do so in time, but there's a bunch of

366
00:37:09,900 --> 00:37:13,980
talks tomorrow, that talk about exactly..
about that thing, so I'm happy to have

367
00:37:13,980 --> 00:37:20,210
this sort of as a lead-in for the day
tomorrow. So what can we.. go, what can we

368
00:37:20,210 --> 00:37:27,090
take from this? The Bundestagswahl 2017 was a
perfect storm for a far-right party like

369
00:37:27,090 --> 00:37:31,570
the AFD. You had a high issue salience of
the topic that is at the center of its

370
00:37:31,570 --> 00:37:42,370
agenda, and you have a sort of unregulated
Wild West of social media. We'll see

371
00:37:42,370 --> 00:37:46,930
how that changes with recent law
changes come into effect, where all of a

372
00:37:46,930 --> 00:37:52,550
sudden the platform itself has some
liability, to which kind of messages are

373
00:37:52,550 --> 00:37:57,310
spread. But if that's effective for
Twitter, is a whole other bag of

374
00:37:57,310 --> 00:38:03,860
worms. So in that sense, that's why I was..
what I was sorting.. hinting at: in this

375
00:38:03,860 --> 00:38:08,640
issue environment, we have people be
interested in the topic that is central

376
00:38:08,640 --> 00:38:16,760
for the party like the AFDs. The media
behaved like pretty surprisingly..

377
00:38:16,760 --> 00:38:22,990
surprisingly predictable?
And did not.. at least for the.. for the

378
00:38:22,990 --> 00:38:30,650
topics, or for the publications, that we
covered, it did so. And for the context

379
00:38:30,650 --> 00:38:35,660
that we're arguing herein, that the AFD
only get like 20 % of the share

380
00:38:35,660 --> 00:38:40,000
towards the end of the campaign, is
something that is a little bit surprising.

381
00:38:40,000 --> 00:38:46,820
And that also leads into a different
question of what does this "Oh it's the

382
00:38:46,820 --> 00:38:54,750
journalists fault!" actually mean? What
does it really mean? This sort of is based

383
00:38:54,750 --> 00:39:00,660
on this normative expectation of the media
being an impartial.. an impartial

384
00:39:00,660 --> 00:39:05,920
deliverer of information and if you think
about what else is going on on the

385
00:39:05,920 --> 00:39:11,540
internet, with alternative media and an
alternative news sphere establishing

386
00:39:11,540 --> 00:39:18,390
itself with news blogs like, well I
don't wanna.. I don't want to call any

387
00:39:18,390 --> 00:39:25,730
names, because.. and so, there's a sort of
scene of far-right fringe blogs in Germany

388
00:39:25,730 --> 00:39:29,600
that we also collected.
And so we're.. further down the line,

389
00:39:29,600 --> 00:39:34,330
we're going to look at what the topics
were, that were covered in that and how

390
00:39:34,330 --> 00:39:39,480
that connected to influencing public
opinion in Germany, but having said this,

391
00:39:39,480 --> 00:39:44,920
with these alternative ways of getting
your news, information being available, if

392
00:39:44,920 --> 00:39:49,930
you have the press, if you have the
mainstream press, not covering a party like

393
00:39:49,930 --> 00:39:56,030
the AFD to a certain extent, you only give
the fodder to those cries of

394
00:39:56,030 --> 00:40:07,260
"Lügenpresse", mendacious press, in
members of the population, that are sort of

395
00:40:07,260 --> 00:40:10,670
at the risk of being lost as audience
members.

396
00:40:10,670 --> 00:40:14,950
So, it's kind of difficult to call the..
to call the shots here and actually point

397
00:40:14,950 --> 00:40:21,000
the fingers at the media, because they
delivered on informing on an interest that

398
00:40:21,000 --> 00:40:30,050
existed in the in the population, before
they reported on something like the AFD.

399
00:40:30,050 --> 00:40:33,520
And with this, I want to leave it at that.
I thank you very much for your attention

400
00:40:33,520 --> 00:40:38,500
and I'm highly, highly eager to hear
questions and prompts and ideas, how we

401
00:40:38,500 --> 00:40:47,726
could pursue this further.
*applause*

402
00:40:47,726 --> 00:40:55,523
Herald: Vielen Dank, thank you very much.
Questions? [unintelligible] any questions?

403
00:40:55,523 --> 00:41:00,480
Feel free to attend the microphones.
Even the microphone I don't

404
00:41:00,480 --> 00:41:05,750
see behind the cameras. Let's start with
number two. *laughs*

405
00:41:05,750 --> 00:41:08,720
Mic 2: [unintelligible] some sound. 
Thank you. Thank you very much for your

406
00:41:08,720 --> 00:41:15,760
amazing work. I've got only one question.
Do you plan on releasing those

407
00:41:15,760 --> 00:41:21,480
collected data and on what license?
Beyer: That's a question that we.. that

408
00:41:21,480 --> 00:41:27,920
we asked ourselves, too. We would love to
collect the data and ultimately it will

409
00:41:27,920 --> 00:41:31,040
happen, but we have to make sure, that we
actually have the right to do so, with the

410
00:41:31,040 --> 00:41:35,770
way we collected it. But we're definitely
looking into that.

411
00:41:35,770 --> 00:41:44,191
Herald: OK, number 5. Yeah, you.
Mic 5: OK. Hello. Is this working? Yeah. It's

412
00:41:44,191 --> 00:41:48,170
tempting - I'm from the Netherlands -
to compare these experiences with the AFD

413
00:41:48,170 --> 00:41:52,270
with the experience in the Netherlands.
You know, we had Wilders, we had Verdonk

414
00:41:52,270 --> 00:41:57,740
we had Fortuyn, now we have Baudet and it
seems that there is a major difference

415
00:41:57,740 --> 00:42:02,360
between.. with the AFD, because presently,
I have frankly, I don't know the name of

416
00:42:02,360 --> 00:42:07,550
the leader of the AFD, it used to be
Frauke Petry and now, I don't know.

417
00:42:07,550 --> 00:42:15,760
But in the Netherlands, the leaders of the..
those populist right-wing parties, they

418
00:42:15,760 --> 00:42:20,990
were.. they were very good in manipulating
the media. They were sending out messages

419
00:42:20,990 --> 00:42:27,540
sustaining a "Köder", in Germany, what's the
word? Like, um, provocating, sending

420
00:42:27,540 --> 00:42:32,311
out provocations and that attracted 
attention of the media. So, there are

421
00:42:32,311 --> 00:42:37,070
people saying that you shouldn't react on
all provocations, but anyway they were

422
00:42:37,070 --> 00:42:43,570
geared to draw attention and I wonder,
whether AFD has been to the same extent

423
00:42:43,570 --> 00:42:52,520
active in the field of drawing attention,
purposely using even agencies that are

424
00:42:52,520 --> 00:42:58,230
specialized in advertising.
Beyer: Great question. There is this idea,

425
00:42:58,230 --> 00:43:05,100
that the AFD was very skillful at sort of
inscenating scandal and purposely doing

426
00:43:05,100 --> 00:43:11,510
things on a public stage that would draw
attention to them. For example, this..

427
00:43:11,510 --> 00:43:17,250
yeah, I say it again.. this
expression by Alexander Gauland, to dispose

428
00:43:17,250 --> 00:43:24,030
of a German politician, or the other
leading candidate Alice Weidel leaving a

429
00:43:24,030 --> 00:43:32,310
talk show, while it was being broadcast. So
there's.. there definitely is this

430
00:43:32,310 --> 00:43:37,351
element of the.. of actually taking a
scandal and using it for your own, for

431
00:43:37,351 --> 00:43:44,980
pushing your own agenda, whereas if they
used ad agencies for their media campaign,

432
00:43:44,980 --> 00:43:53,270
they did, their campaigning was highly
professionalized, in terms of what their

433
00:43:53,270 --> 00:43:57,370
posters were and how their campaign ads
were worked.

434
00:43:57,370 --> 00:44:05,400
And they did work with a company, that also
was involved with Donald Trump's campaign.

435
00:44:05,400 --> 00:44:12,680
But in terms of.. sort of new media or
like online media – it's not that new

436
00:44:12,680 --> 00:44:19,350
anymore – and in terms of what they did on
online media, I.. I only have an

437
00:44:19,350 --> 00:44:28,570
anecdotal sense, if they use something like
bots, which is also a way of buying,

438
00:44:28,570 --> 00:44:34,760
buying attention. I can.. I can sort
of tell you about one specific case, where

439
00:44:34,760 --> 00:44:42,220
we investigated, which Twitter users were
the most active in tweeting on the AFD on

440
00:44:42,220 --> 00:44:47,370
German Twitter – tomorrow's a talk about a
Twitter user called Ballerina, which is

441
00:44:47,370 --> 00:44:50,460
a name that has been out there which..
there's great education, that that is

442
00:44:50,460 --> 00:44:54,640
definitely a bot, that has been planted
and has been controlled by someone else or

443
00:44:54,640 --> 00:45:06,000
by sort of.. by any group of
actors that is not actually a ballerina.

444
00:45:06,000 --> 00:45:14,210
What we found was a Twitter user called
Teletubbies007, that tweeted in those three

445
00:45:14,210 --> 00:45:22,900
weeks, that we surveyed, 6.500 times and
mostly just retweeted, retweeted calls to

446
00:45:22,900 --> 00:45:29,990
go and cast your ballot, that were all put
out by the central AFD accounts. And it

447
00:45:29,990 --> 00:45:33,710
didn't have a lot of followers, like
something 500 or so, but it just kept

448
00:45:33,710 --> 00:45:39,850
retweeting over and over and over and over.
And when we actually wanted to check out

449
00:45:39,850 --> 00:45:46,360
the page of that bot, it was deleted,
the user was deleted.

450
00:45:46,360 --> 00:45:51,140
So there's, to answer your question, um,
this high.. this degree of

451
00:45:51,140 --> 00:46:00,040
personalization that the Partij voor
de Vrijheid has in the Netherlands is not

452
00:46:00,040 --> 00:46:03,180
as extreme for the AFD in Germany, because
there's more leading candidates and

453
00:46:03,180 --> 00:46:09,300
there's internal rifts like Geerd Wilders
is basically his own party. That's not the

454
00:46:09,300 --> 00:46:15,240
same. But the strategy to use scandal and
to use something that is outrageous and

455
00:46:15,240 --> 00:46:18,970
push the boundaries a little bit more,
then jump back and say "Oh no, we did not

456
00:46:18,970 --> 00:46:24,260
mean that at all in this way", that is
the exact same spot on strategy they used.

457
00:46:24,260 --> 00:46:27,630
Mic 5: Perhaps I should add that Wilders
made it like..

458
00:46:27,630 --> 00:46:31,530
Herald: Excuse me, many people queuing.
Mic 5: Okay. Then I'll stop.

459
00:46:31,530 --> 00:46:35,020
Herald: Okay, thank you. We have questions
from the internet, then.

460
00:46:35,020 --> 00:46:40,660
Signal Angel: Yes. <QgeeTG>(?) is asking: "Why
did you come to the conclusion that this

461
00:46:40,660 --> 00:46:44,620
was a special election, while the last
election in Austria has exactly the same

462
00:46:44,620 --> 00:46:49,350
issues? Don't you see this as some sort of
an global effect?"

463
00:46:49,350 --> 00:46:57,350
Beyer: That's true, a Syrian civil war
that pushes people to flee from,

464
00:46:57,350 --> 00:47:00,460
from war and save their livelihood, is
something that is not only felt in

465
00:47:00,460 --> 00:47:05,050
Germany, but for the context of Germany,
it's a special election. That's.. this sort

466
00:47:05,050 --> 00:47:11,330
of situations never.. has never occurred in
this way before. But absolutely, each

467
00:47:11,330 --> 00:47:20,740
election in Europe basically since 2015
was a special election in that sense. But

468
00:47:20,740 --> 00:47:25,470
not in terms of the outcomes, in a way,
that.. because far-right parties in other

469
00:47:25,470 --> 00:47:30,500
European countries already had, had their
foot in the door and especially in Austria,

470
00:47:30,500 --> 00:47:35,550
where.. with the FPÖ were pretty well
established with previously having been

471
00:47:35,550 --> 00:47:39,801
part of a government. And now
being part of the government again. But

472
00:47:39,801 --> 00:47:45,800
for Germany, in what the issues were, that
were top of people's minds: that's the

473
00:47:45,800 --> 00:47:51,370
special case that I meant.
Herald: OK, microphone number 3, please.

474
00:47:51,370 --> 00:47:58,120
Mic 3: Thank you, first I really
appreciate the sincerity and transparency

475
00:47:58,120 --> 00:48:03,010
of your talk, thank you very much, we need
more of this in such circumstances and

476
00:48:03,010 --> 00:48:10,660
maybe less polemics sometimes. There's
just a little trifle in your method, where

477
00:48:10,660 --> 00:48:18,190
I was wondering: how did you filter the
"Linke" and "Grüne" stuff. Did you..

478
00:48:18,190 --> 00:48:23,760
yeah, how exactly did you do it? Did you
maybe count all the mentions of "Grün"

479
00:48:23,760 --> 00:48:29,840
with a capital and non-capital "G", and
"Linke" with a capital L and non-capital

480
00:48:29,840 --> 00:48:34,890
and then filter it out further? Or did you
do it the other way around? I know, that

481
00:48:34,890 --> 00:48:41,310
you focused specifically on the AFD stuff,
and maybe you were focused on representing

482
00:48:41,310 --> 00:48:46,570
all the parties that might be relevant.
But I would still be interested in that

483
00:48:46,570 --> 00:48:50,850
part, thanks.
Beyer: That's a great question. The thing

484
00:48:50,850 --> 00:48:55,450
is, that we used.. when we actually put all
that.. when after we collected text, before

485
00:48:55,450 --> 00:49:01,510
we put it through the unloading
methods, we put it all into lowercase.

486
00:49:01,510 --> 00:49:09,581
Just so we could have a consistent way of
analyzing. And with capitalization, it's

487
00:49:09,581 --> 00:49:16,290
kind of.. sometimes it just trips up the
way to treat this. And that's why you

488
00:49:16,290 --> 00:49:21,370
ran into these issues with "Linke und
Grüne" where we had to resort to only

489
00:49:21,370 --> 00:49:25,510
taking basically the candidates names and
then also "Linke Partei" and "Grüne

490
00:49:25,510 --> 00:49:35,750
Partei" and a few conjugations, so "der Linken
Partei", "den".. right, like grammatically..

491
00:49:35,750 --> 00:49:40,500
the cases, we only, like, we conjugated
them through. Yeah, but we.. since our focus

492
00:49:40,500 --> 00:49:43,910
was on the AfD, we weren't especially
concerned with that, which is

493
00:49:43,910 --> 00:49:48,350
unfortunate, I admit that, but for the
purpose of this talk, we decided to just

494
00:49:48,350 --> 00:49:53,500
use this workaround.
Mic 3: OK, thanks.

495
00:49:53,500 --> 00:49:56,260
Herald: OK. Microphone six,
please.

496
00:49:56,260 --> 00:49:59,130
Mic 6: Hello, thanks for your interesting
presentation.

497
00:49:59,130 --> 00:50:03,480
I'm wondering if you and your team..
so, you'd.. you looked at mentions

498
00:50:03,480 --> 00:50:07,760
of the different parties, but I'm wondering
if you looked at the content of the articles

499
00:50:07,760 --> 00:50:12,650
and how they talked about it, if they were
talked about positively or negatively.

500
00:50:12,650 --> 00:50:15,110
Beyer: Thank you very much, that's a great

501
00:50:15,110 --> 00:50:19,070
question, that we actually did consider.
And I'll answer this question with a

502
00:50:19,070 --> 00:50:23,520
counter question, as social scientists
like to do. Anyone in this room use Amazon

503
00:50:23,520 --> 00:50:28,591
Mechanical Turk and works on hits to earn
a few cents here and there? No? OK, so I

504
00:50:28,591 --> 00:50:37,580
can speak freely. There's a.. there's a
method that uses cheap labor on Amazon

505
00:50:37,580 --> 00:50:42,920
Mechanical Turk and presents each worker
with two sentences, out of which they have

506
00:50:42,920 --> 00:50:50,000
to change the one that is more positive.
And so we wanted to use this to train a

507
00:50:50,000 --> 00:50:55,120
machine learning algorithm to actually get
a way to gauge the sentiment of positive

508
00:50:55,120 --> 00:50:58,930
and negative in the text that we had
collected. We started that in early

509
00:50:58,930 --> 00:51:06,730
December and we had a, like, a workbook with
4.000 so-called hits, 4.000 little jobs,

510
00:51:06,730 --> 00:51:16,140
4.000 comparisons and when this job was
done, five or six days later, we sort of put

511
00:51:16,140 --> 00:51:23,490
that through a test and compared it with
our own hand-coding that we had done.

512
00:51:23,490 --> 00:51:31,560
And it turned out that one worker on Amazon
Mechanical Turk spent over seven hours and

513
00:51:31,560 --> 00:51:38,510
worked.. of those 4.000 little jobs
that we had, he worked 3.980.

514
00:51:38,510 --> 00:51:45,250
And over 1.400 of which 
he did in less than two seconds.

515
00:51:45,250 --> 00:51:53,840
Which is unfortunate, because: a) this
person.. so, this person - right, "Person?, Question

516
00:51:53,840 --> 00:51:59,870
Mark" - probably used a script, probably
used a bot or just randomly clicked. The

517
00:51:59,870 --> 00:52:06,770
coding didn't match up at all with what we
did hand-wise ourselves and that really

518
00:52:06,770 --> 00:52:13,710
screwed up our approach there. If any of
you plan on doing some hits in the new

519
00:52:13,710 --> 00:52:18,860
year for Amazon Mechanical Turk and you're
asked to compare two sentences that

520
00:52:18,860 --> 00:52:23,470
mention a political actor in Germany, you
can send me an email and maybe a

521
00:52:23,470 --> 00:52:28,110
screenshot and tell me how much you
appreciate that we're paying six cents for

522
00:52:28,110 --> 00:52:34,440
each comparison. But that's the story, why
we haven't.. we don't have any sentiment in

523
00:52:34,440 --> 00:52:40,260
this analysis here.
Herald: [unintelligible]

524
00:52:40,260 --> 00:52:44,840
Mic ?: Hello. I'm from Denmark, so in this
context, I'm very much a ghost of

525
00:52:44,840 --> 00:52:47,500
Christmas future. 
*Beyer laughs*

526
00:52:47,500 --> 00:52:49,040
Mic ?: In your Twitter data,

527
00:52:49,040 --> 00:52:54,660
where you take Retweets as well, do you
determine what are quotes and what

528
00:52:54,660 --> 00:53:00,290
are direct Retweets? Because in my
experience, and I work with this in Denmark

529
00:53:00,290 --> 00:53:07,360
and in the UK, a lot of people like to
distance themselves from what the AfD and

530
00:53:07,360 --> 00:53:15,740
similar are saying by quoting everything
they're saying and giving them the press.

531
00:53:15,740 --> 00:53:21,360
Beyer: That's a very good point to make.
We did not make any distinction

532
00:53:21,360 --> 00:53:27,020
between quotes and Retweets, but we did
filter, based on 5 Retweets, by thinking:

533
00:53:27,020 --> 00:53:31,280
OK, if you occasionally feel like you
have to point something out that is

534
00:53:31,280 --> 00:53:36,840
outrageous and ridiculous, that a person,
a member of a party, says on Twitter, you

535
00:53:36,840 --> 00:53:41,100
would be inclined to do so less than a
certain amount of time. We also tried it

536
00:53:41,100 --> 00:53:45,760
with other cut-offs. The graph basically
always looked the same. But if we think

537
00:53:45,760 --> 00:53:52,700
about what this means for how the demand-
side is influenced, it doesn't matter.

538
00:53:52,700 --> 00:53:57,210
Basically, if you're retweeting out of
endorsement or out of ... out of ...

539
00:53:57,210 --> 00:53:59,420
Mic ?: Spite.
Beyer: ... out of spite, that's right.

540
00:53:59,420 --> 00:54:03,630
That's the logic, why we decided to use
mentions and Retweets.

541
00:54:03,630 --> 00:54:05,240
Mic ?: Thank you.
Herald: Another question

542
00:54:05,240 --> 00:54:09,180
from the internet?
Signal: Yes. Luke23 is asking: Do you

543
00:54:09,180 --> 00:54:12,960
think that the window of commonly
acceptable ideas, the so-called Overton

544
00:54:12,960 --> 00:54:21,500
window, was shifted to the right by the
ideas of the AfD echoed in the media?

545
00:54:21,500 --> 00:54:25,890
Beyer: That's a good question. That's a
good question. Something that comes to

546
00:54:25,890 --> 00:54:30,700
mind here, is that media use is
epiphenomenal - you're sort of

547
00:54:30,700 --> 00:54:38,130
likely.. but the question is, like: Do you
think.. does something happen in you,

548
00:54:38,130 --> 00:54:42,250
because you use a certain media outlet, or
do you use a certain media outlet, because

549
00:54:42,250 --> 00:54:48,000
something happened in you ?
From the sense that I got, I would say that

550
00:54:48,000 --> 00:54:53,530
the degree to what is.. what is acceptable,
definitely was shifted over the course of

551
00:54:53,530 --> 00:54:57,740
this campaign, that all of a sudden we're
questioning, if remembering the Holocaust

552
00:54:57,740 --> 00:55:03,770
should be something that is at the heart
or very close to German identity.

553
00:55:03,770 --> 00:55:08,810
That's something that a political
scientist would have never expected, that

554
00:55:08,810 --> 00:55:15,891
this cleavage can be opened up again in a
way that is so potent as it did now. So it

555
00:55:15,891 --> 00:55:22,790
definitely did something to the overall
discourse in Germany. Whereas that is an

556
00:55:22,790 --> 00:55:30,580
effect of media reporting on the AFD,
would require us to use something like

557
00:55:30,580 --> 00:55:35,930
this.. the sentiment analysis, to actually
determine how the media talked about which

558
00:55:35,930 --> 00:55:42,100
aspect of the AFD agendas.

559
00:55:42,100 --> 00:55:47,090
Herald: I can see some movement behind
microphone number 8. I'm sorry. *laughs*

560
00:55:47,090 --> 00:55:51,910
Mic 8: Thank you very much. Thank you for
your work, I still do have a critical

561
00:55:51,910 --> 00:55:57,930
question. Basically, the things you showed
is something like we all know, yeah? We

562
00:55:57,930 --> 00:56:03,030
could see this happening last year, and so -
I mean this year, in the last election. So

563
00:56:03,030 --> 00:56:08,710
I am wondering now, whether the method you
used, which was basically focusing on

564
00:56:08,710 --> 00:56:15,820
quantity, is in a sort of mirroring what
was happening. And I'm wondering if you

565
00:56:15,820 --> 00:56:22,340
would work.. keep working on it. Like, you
used buzzwords and you used "the media"

566
00:56:22,340 --> 00:56:31,000
instead of, like, narrowing it down, or more..
using more specific questions and I was

567
00:56:31,000 --> 00:56:37,320
wondering, if you have these results now and
you have proof for them? What are your next

568
00:56:37,320 --> 00:56:44,270
questions and how can you continue to use
these.. the data you have, to make it more

569
00:56:44,270 --> 00:56:50,500
specific, so we can really have some outcome
and some conclusions coming from this?

570
00:56:50,500 --> 00:56:52,870
Beyer: It's a absolutely wonderful
question.

571
00:56:52,870 --> 00:56:58,021
Of course, we thought about using
this data further down the line. We.. our

572
00:56:58,021 --> 00:57:04,370
initial plan was, to connect this not just
with salience data that we derive from

573
00:57:04,370 --> 00:57:07,710
Google searches.
We also have Facebook data that we

574
00:57:07,710 --> 00:57:11,980
collected, that we wanted to look into,
but there.. it's a bit challenging, to

575
00:57:11,980 --> 00:57:19,780
actually analyze comments in depth onto
language, because language tends to be way

576
00:57:19,780 --> 00:57:26,920
more fluid and you have certain problems
with selection and self-selection. So you

577
00:57:26,920 --> 00:57:30,480
really, really have to be careful to cross-
connect, which person that comments on

578
00:57:30,480 --> 00:57:38,210
Facebook is the same person and thus, if
you only do quantitative stuff, would

579
00:57:38,210 --> 00:57:44,210
appear disproportionally. As I mentioned,
we have also collected data from far-right

580
00:57:44,210 --> 00:57:52,490
blogs, from "news" blogs, that very
actively endorsed the AFD and their topics

581
00:57:52,490 --> 00:57:59,110
and so we're planning to pull this into
the analysis along with data from the

582
00:57:59,110 --> 00:58:04,170
German Longitudinal Election Study, where
in this time frame, that we surveyed, in the

583
00:58:04,170 --> 00:58:10,800
data, each day 100 people in Germany were
called up and asked about their feelings

584
00:58:10,800 --> 00:58:16,250
toward specific parties and actors. So we
actually have day-by-day data, once it

585
00:58:16,250 --> 00:58:21,730
comes out, on how people.. what people
thought about those actors. So we're

586
00:58:21,730 --> 00:58:28,020
planning to pull that in, as a more
reliable measure for salience.

587
00:58:28,020 --> 00:58:30,790
Herald: Thank you very much. I'm very
sorry, but time's up, so there will be no

588
00:58:30,790 --> 00:58:35,030
more questions right now in front of the
audience. Alexander Beyer, thank you very

589
00:58:35,030 --> 00:58:38,030
much. A warm applause, please.
*applause*

590
00:58:38,030 --> 00:58:40,864
Beyer: Thank you.
*applause continues*

591
00:58:40,864 --> 00:59:01,744
*postroll music*