=Paper=
{{Paper
|id=Vol-2111/paper4
|storemode=property
|title=What Does it Mean to be a Wutbürger? A First Exploration
|pdfUrl=https://ceur-ws.org/Vol-2111/paper4.pdf
|volume=Vol-2111
|authors=Manfred Klenner
|dblpUrl=https://dblp.org/rec/conf/esws/Klenner18
}}
==What Does it Mean to be a Wutbürger? A First Exploration==
https://ceur-ws.org/Vol-2111/paper4.pdf
What does it mean to be a Wutbürger?
A first exploration.
Manfred Klenner
Institute of Computational Linguistics
Andreasstrasse 15, 8050 Zurich, Switzerland
klenner@cl.uzh.ch
Abstract. In this paper, we undertake an attempt to characterize the
world view of what are called Wutbürger in Germany, that is citizen who
are enraged by the current political and social situation. In order to find
out what makes a Wutbürger a Wutbürger, we analyze Facebook posts
on the basis of a lexical resource where nouns, adjectives and verbs are
classified according to Plutchik’s primary emotions. We also introduce
new polar roles of verbs that help to identify the writer perspective. This
way, we are able to identify targets and the Wutbürger’s stance towards
them. As textual data, we utilize about 100,000 Facebook posts of a
German right-wing party whose members are obvious exemplars of the
notion of a Wutbürger.
1 Introduction
In a number of western societies populism has (re)entered the scene and espe-
cially rampages in the social media. Hate speech, shit storms etc. are extreme
forms of such undemocratic tendencies. In Germany, the notion of a Wutbürger
has been coined, that is, citizen who are disappointed by the government and the
social situation and their (verbal) behavior seem to be driven by rage (German
Wut). A new, right-wing party evolved, the AfD (Alternative für Deutschland).
We have access to about 100,000 Facebook posts of the AfD including reader
comments that mostly stem from AfD proponents - who clearly form a sub-
set of German Wutbürger. Our research question was: Can we find out, how a
Wutbürger perceives the world and what, after all, is the objective of his Wut.
A first step towards this goal is to measure the emotional fingerprint of the
texts produced by Wutbürger and compare it to the fingerprint of a related text
genre. We use the Tübinger (German newspaper) Treebank (TüBa-D/Z) [11]
as a reference corpus. In order to compare the emotional load of the AfD texts
(303,563 sentences) and the TüBa-D/Z texts (95,595 sentences), we perform a
lexicon-based analysis. That is, we count the primary emotions by the use of
words that are indicative of these emotions. This is a straightforward approach,
but it should tell us reliably what the prevalent emotions in these texts are and
whether the AfD texts are more loaded than the newspaper reference texts.
The emotional fingerprint does not tell us anything about the world view of
a Wutbürger: who are his proponents and who are his opponents? We would like
What does it mean to be a Wutbuerger 33
to exploit the idea that the identification of these targets can be supported and
accomplished by a more fine-grained classification of lexical items. We not only
assign primary emotions to words (verbs, nouns and adjectives), we also identify
those words that have an implicit writer perspective and explicate which one it
is. For instance, the adjective ineffable in a phrase like the ineffable chancellor
expresses that the writer has a negative attitude towards the referent of the noun
and a sentence like Merkel jerks the German citizen around allows to infer that
the writer believes that the referent at subject position is an immoral actor, a
cheater one might say. Also, the direct object is perceived as a victim of the
cheater. Since the writer is against the cheater, he is in favor of the victim.
2 Lexical Resources
Starting with the freely available lexicons described in [2]1 and [4]2 we identified
those verbs, nouns and adjectives that have an emotional dimension (e.g. to
love, to hate, gratitude, joy, pleasant, happy). We then classified each of the 168
verbs, 225 nouns and 300 adjectives according to Plutchik’s [7] eight primary
emotions which are anger, fear, sadness, disgust, surprise, anticipation, trust,
and joy. This was done by two annotators, who achieved a Kappa value of κ
0.73. Furthermore, we annotated those verbs, nouns and adjectives that refer to
moral, e.g. to lie, donation. See Figure 1 for an overview (pos, neg are shortcuts
for positive, negative respectively). Kappa was κ = 0.66.
pos emotion neg emotion pos moral neg moral pos factual neg factual #
verb 49 119 15 71 553 1170 1977
adj 118 182 286 569 1010 1103 3268
noun 91 134 104 436 663 1229 2657
Fig. 1. Lexicon Overview: Word Frequencies
The columns for factual denote words that are positive or negative without
reference to either (a particular) emotion or moral. We could say that they are
positive or negative on a factual level. For instance to sicken, recover, congratu-
late are examples of such verbs, whereas mistake, disease, transparency, security,
right, wrong are examples for such nouns and adjectives. This is a crucial dis-
tinction: such words do not indicate a writer perspective, but the contribute to
polarity decisions, nevertheless.
We took the 254 verbs classified as either belonging to the emotion or moral
dimension as a basis for further annotations. We identified 58 verbs with a very
strong writer perspective either on the actor or the experiencer role or on both
(to cheat, to jerk sb around ). We then coined for the six verb classes derived that
way special role labels. The set of agent roles is: prole, baiter, hater, torturer,
hypocrite, choleric. Experiencer roles are sufferer and victim. To give an exam-
ple: the verb flare up (aufbrausen) bears the emotion anger and the semantic
1
http://bics.sentimental.li/files/8614/2462/8150/german.lex
2
https://pub.cl.uzh.ch/projects/opinion/lrec data.txt
34 Manfred Klenner
role of the subject is that of choleric. Our hypothesis is that these roles better
capture the writer perspective, since they express how the writer conceptualizes
these referents. Note that we assign these roles to subcategorization frames, not
to verbs. We specified these verbs along the line proposed by [4]. That is, we
modeled the various subcategorization frames of a verb and assigned it a polar
effect (positive or negative) and for some of the verbs also a dedicated polar role
(sufferer, torturer etc.). We thus were able to find out who the AfD believes to
be a torturer, a baiter etc. and who suffers from the situation described.
3 Corpus Statistics and Lexical Coverage
Our present endeavor is basically one that exploits an existing, but carefully
refined and augmented lexical resource. Especially our new verb classes with a
new kind of polar roles are meant to make the writer perspective more nuanced.
We are at the very beginning of a sophisticated study. At the moment, however,
there is no gold standard and thus no machine learning involved.
We found 9,012 verb types in the Facebook posts, which gave us altogether
419,034 verb tokens in 303,563 sentences (word tokens altogether: 5,249,613).
The 1052 verb types of our lexicon found in the data (61 verbs did not occur),
amounts to 83,658 verb tokens, which is - taking into account that one sentence
might have more than one model verb - about 25% coverage (a model verb in
each 4th sentence). Our verb resource seems to have a good coverage, thus. If
we just look at moral verbs, we get 11,153 hits, 64 of the 85 moral verb types do
occur in the posts. The 168 emotional verbs occur with a frequency of 17,102.
In order to quantify the emotional load of the Facebook posts, we used the
Tübinger Treebank (TüBa-D/Z) as a reference corpus. The TüBa-D/Z comprises
95,595 sentences. The coverage of our verb resource is again quite good: we found
930 verb types with 22,679 verb tokens, which is a coverage of 23.79% (again
almost each 4th sentence bears a model verb).
4 Emotional Fingerprint
We use our emotion lexicon in order to diagnose the emotional load of the AfD
posts. In Fig. 2, we compare the AfD posts and the newspaper text wrt. to
the emotions present. We determined the frequency of words belonging to a
particular emotion and normalized by the total number of emotion words (found
in the posts) (left hand side), and by the number of sentences (right hand side),
respectively.
As we can see from Fig. 2, fear is the most prominent emotion of a Wutbürger
and not anger (a prestage of rage) while at the same time sadness is not a
prevalent emotion of a Wutbürger. All other emotions are almost identically
distributed in both, AfD posts and newspaper text. Our expectation, namely
that the AfD posts would have a higher emotional load than news texts, was not
confirmed.
What does it mean to be a Wutbuerger 35
Fig. 2. Emotional Fingerprints of AfD Facebook Post and a Newspaper Corpus
We also had a look at the moral dimension. The TüBa-D/Z refers to 7490
nouns and adjectives classified as positive (32%) or negative (60%) from a moral
perspective, i.e. 7.8% of the sentences refer to that dimension. In the AfD texts,
32167 tokens were found, which is about 10.6% (73% negative, 27% positive).
This clearly shows that (negative) moral argumentation is a central attitude of
a Wutbürger. If we have a look at verbs the picture is similar: 1.6% of the news
texts contain a moralizing verb, whereas 2.3% of the AfD posts do so.
5 Target Identification and Stance Analysis
A polar role is the label for the logical subject (agent) or object (theme, patient,
experiencer) that indicates the positive or negative role its filler plays. The in-
ventory of polar roles is not fixed, yet. We have defined a couple of fine-grained
polar roles that are meant to indicate a more nuanced writer perspective. These
roles are baiter, hater, choleric, hypocrite, prole, torturer and sufferer, victim.
The definition of these roles is straightforward: we just had to fix the corre-
sponding verbs and determine which semantic role bears which polar role. Take
the polar role prole. There is a number of verbs in German (we have identified
18) that indicate that the writer implicitly classifies the agent of such a verb
as a prole (anlabern (to chat so up), anpöbeln (to accost sb)). Thus, the agents
of such verbs are negative targets from the point of view of the writer. It turns
out that in the AfD texts journalists, do-gooder, politicians, asylum seekers, the
print media are, among others, conceptualized as proles.
In order to derive these writer perspectives, we have parsed the AfD posts
with a dependency parser [10], normalized the parse trees (e.g. passive voice) and
extracted the filler of the polar roles. This way we found e.g. that (the targets)
Merkel, the German government, the police and the press are baiter, hater and
torturers etc. The AfD, the German citizen and women were victims and sufferer.
Clearly, these lists are not perfect. There are (third person) pronouns in it and
also words denoting non-actors. The goal of this explorative study was to get a
36 Manfred Klenner
proof of concept not a full-fledged evaluation. Nevertheless, we have carried out
a small evaluation in order to find out where the noise comes from. We randomly
took 50 sentences where the German chancellor Angela Merkel was the logical
subject of a morally negative verb (e.g. cheat, threaten, diss, violate ) and 50
where the AfD was the experiencer or patient of such a verb. Only 9 out of the
100 decisions were wrong due to 4 parsing errors and 5 modal constructions that
erroneously passed our modal filter.
An interesting finding concerns the role of the AfD (i.e. Wutbürger) itself.
If we look at those who are hated, we get: Arabs, strangers, Merkel, Muslims,
comrads but also Germany and the AfD. A closer inspections reveals that the
Wutbürger do not disguise or veil their rage. They use verbs with AfD (or I or
we) as agents that indicate that they are haters.
Our verb resource also allows for more sophisticated inferences. We have
coined the notion of a violator of morality for the following set of actors: the set
of actors of a verb that casts a negative effect on its object which is - according
to the polarity lexicon - positive:
λX.(∃Verb, Y: subj(Verb,X)∧effect(Verb,obj,neg)∧obj(V erb, Y )∧polarity(Y, pos))
An example is Merkel destroys the security of Germany where security is
positive and destroy casts a negative effect on the direct object (obj)
If we, however, change polarity(Y,pos) to polarity(Y,neg) than we get a strong
proponent of the AfD: to disapprove something negative is positive.
6 Related Work
One topic of this paper is lexicon-based, document-level emotion detection. For
an overview of similar approaches see e.g. [1]. We have specified the first German
emotion lexicon, where words are associated with primary emotions and - if
applicable - writer perspectives.
The role verbs play in sentiment analysis and stance detection has received
increased attention over the last years, cf. [6], [9], [3], [8], [5]. The main difference
to our German verb resource is that we not only specify the polar effects (positive
or negative) a verb casts on its semantic roles, we also strive to assign fine-
grained role labels such as torturer etc. Again this is meant to allow for a finer
nuanced writer perspective, which not only helps to identify targets, but also
the stance taken towards those targets. Another distinctive feature is that we
combine bottom-up and top-down information in order to derive stance (see last
section).
7 Conclusions
We have introduced two new resources for German: a fine-grained verb resource
with polar roles that reveal the writer perspective, and an emotion lexicon where
words are classified as one of eight primary emotions. Also words related to moral
are specified. We used this in order to find out whether Wutbürger texts do have a
What does it mean to be a Wutbuerger 37
clear emotional fingerprint compared to news texts. We found that fear (and not
Wut, i.e. rage) is the prevalent emotion and that Wutbürger significantly more
often argue on the basis of moral than a reference newspaper corpus. Another
insight is that Wutbürger do not hide their rage (in their own sentences they often
occupy negative polar roles such as hater). More sophisticated search pattern on
the basis of top down and bottom up restrictions give rise to interesting inferences
(someone who disapproves something positive is a violator of morality). All this
is meant as a first explorative study: is our lexicon large enough to be useful, is
our fine-grained verb resource broadly applicable. A fuller answer to the question
raised in the title must await a thorough empirical investigation.
References
1. Haji Binali, Chen Wu, and Vidyasagar Potdar. Computational approaches for
emotion detection from text. In Proceedings of IEEE Intern. Conf. on Digital
Ecosystems and Technologies, 2010.
2. Simon Clematide and Manfred Klenner. Evaluation and extension of a polarity
lexicon for German. In Proceedings of the First Workshop on Computational Ap-
proaches to Subjectivity and Sentiment Analysis (WASSA), pages 7–13, 2010.
3. Lingjia Deng and Janyce Wiebe. Joint prediction for entity/event-level sentiment
analysis using probabilistic soft logic models. In Proceedings of the 2015 Confer-
ence on Empirical Methods in Natural Language Processing, EMNLP 2015, Lisbon,
Portugal, September 17-21, 2015, pages 179–189, 2015.
4. Manfred Klenner and Michael Amsler. Sentiframes: A resource for verb-centered
German sentiment inference. In Proceedings of the Tenth International Conference
on Language Resources and Evaluation (LREC 2016), Paris, France, may 2016.
European Language Resources Association (ELRA).
5. Manfred Klenner, Don Tuggener, and Simon Clematide. Stance detection in Face-
book posts of a German right-wing party. In LSDSem 2017/LSD-Sem Linking
Models of Lexical, Sentential and Discourse-level Semantics, April 2017.
6. Alena Neviarouskaya, Helmut Prendinger, and Mitsuru Ishizuka. Semantically
distinct verb classes involved in sentiment analysis. In IADIS AC (1), pages 27–
35, 2009.
7. Robert Plutchik. A general psychoevolutionary theory of emotion. Academic press,
NewYork, 1980.
8. Hannah Rashkin, Sameer Singh, and Yejin Choi. Connotation frames: A data-
driven investigation. In Proceedings of the 54th Annual Meeting of the Association
for Computational Linguistics (ACL), Berlin, Germany, Angust 2016.
9. Kevin Reschke and Pranav Anand. Extracting contextual evaluativity. In Proc. of
the Ninth International Conf. on Computational Semantics, pages 370–374, 2011.
10. Rico Sennrich, Martin Volk, and Gerold Schneider. Exploiting synergies between
open resources for German dependency parsing, POS-tagging, and morphological
analysis. In Recent Advances in Natural Language Processing (RANLP 2013),
pages 601–609, September 2013.
11. Heike Telljohann, Erhard W. Hinrichs, Sandra Kübler, Heike Zinsmeister, and
Kathrin Beck. Stylebook for the Tübingen treebank of written German. Technical
report, Universität Tübingen, Seminar für Sprachwissenschaft, 2009.