ArticlePDF Available

Towards Balance and Boundaries in Public Discourse: Expressing and Perceiving Online Hate Speech (XPEROHS)


Abstract and Figures

This study presents an overview and preliminary findings from the XPEROHS-project on hate speech in online contexts. The data is extracted from large-scale Facebook and Twitter corpora, while comparing linguistic instantiations of hate speech in the Danish and German languages. Findings are based on four sub-projects involving the semantics and pragmatics of denigration, the covert dynamics of hate speech, perceptions of spoken and written hate speech, and rhetorical hate speech strategies employed in online interaction. The results demonstrate both overt and covert hate speech towards minority groups, especially Muslims, that are symptomatic of larger societal othering processes and stigmatization. KEYWORDS hate speech, ethnophaulisms/ethnic slur terms and denigration, social media, corpus linguistics, German/Danish language, immigration discourse'
Content may be subject to copyright.
Towards Balance and Boundaries in Public
Discourse: Expressing and Perceiving Online
Hate Speech (XPEROHS)
1School of Languages and Cultures, University of Sheeld,
2Department of Language and Communication, University of Southern Denmark
3Mads Clausen Institute, University of Southern Denmark
4Department for the Study of Culture, University of Southern Denmark
1. Introduction
Hate speech is a growing source of concern. Particularly in online contexts, increased inci-
dences of hate speech involving ethnicity, nationality, and religion have been observed (Fox-
man & Wolf 2013). Yet, the very notion of hate speech remains highly controversial; there is a
lack of consensus about its denition and impact, while the motivation and justication for its
criminalisation and reg ulation are inexorably caught between the need to protect human rights
of equality a nd dignity and the civ il liberty of freedom of expression (Herz & Mol nar 2012). Con-
sidering the pressure that hate speech exerts on the pillars of modern civilization, it is striking
how little is know n about the linguist ic and communicative mechanisms u nderlyin g the expres-
sion and perception of hate speech. This gap applies, in particular, to written online communi-
cation on media platforms such as Twitter or Facebook, but also to actual spoken communica-
This study presents an overview and preliminary ndings from the XPEROHS-project on
hate speech in online contex ts. The data is ex tracted from large-scale Facebook and Twitter
corpora, while comparing linguistic instantiations of hate speech in the Danish and German
languages. Findings are based on four sub-projects involving the semantics and pragmatics
of denigration, the covert dynamics of hate speech, perceptions of spoken and written hate
speech, and rhetorical hate speech strategies employed in online interaction. The results
demonstrate both overt and covert hate speech towards minority groups, especially Muslims,
that are symptomatic of larger societal othering processes and stigmatization.
hate speech, ethnophaulisms/ethnic slur terms and denigration, social media, corpus
linguistics, German/Danish language, immigration discourse'
Baumgarten et al.
tion in everyday oine interaction (Assimakopoulos et al. 2017). Even less is known about how
the mechanisms underlying the expression of hate speech are perceived by ordinary language
users; do they operate in similar ways in both written and spoken language? Can speakers in
oral communication always argue that everything was not meant seriously and literally? Or is it
possible to dene acoustic indicators that unmask hate speech reliably?
The XPEROHS project, funded by the Velux Foundation (project no. 95-16416), aims to ll
some of these gaps for the Danish and German languages; the project is divided into four inter-
connected sub-projects employing radically empirical approaches which address hate speech
from the perspectives of production and perception. Sub-project 1 focuses on the use and per-
ception of slurs, dehumanising metaphors and metonyms in Danish and German. Sub-projects
2 (for Danish) and 3 (for German) are concerned with a wider range of subtle mechanisms for
expressing hate speech; t hey focus on morphological, syntactic and discourse level phenomena
and include large-scale corpus analyses of Danish and German as well as a smaller ‘case study’
investigation of rhetorical strategies in the comment sections of two Danish news providers.
Sub-project 4 uses perceptual experiments to investigate how various social groups judge, and
are aected by, expressions of hate speech.
While denitions of hate speech are subject to specic and more locally dened cultural no-
tions, the working denition adopted here is from the European Commission against Racism
and Intolerance in its General policy recommendation No. 15 (2015).
[T]he advocacy, promotion or incitement, in any form, of the denigration, hatred
or vilication of a person or group of persons, as well as any harassment, insult, ne-
gative stereotyping, stigmatization or threat in respect of such a person or group of
persons and the justication of all the preceding t ypes of expression on the grounds
of ‘race’, colour, descent, national or eth nic origin, age, disabilit y, langu age, religion
or belief, sex, gender, gender identity, sexual orientation and other personal charac-
teristics or status. (COE 2015)
We have selected this broad denition as it captures the heterogeneity of the hate speech con-
cept, which Brown (2017) argues does not exhibit the simple compositional semantics of hate +
speech (cf. Perry 2005); it is best understood in terms of Wittgensteinian family resemblances,
encompassing a multiplicity of meanings and forms of expression: thus Haas (2012: 130) sees
hate speech, stereotypical talk, and prejudiced communication as a “family of concepts”. The
use of a working denition, however, does not exclude sensitivity towards the socio- cultural na-
ture of hate speech as a product of, and a practice embedded within, specic historical, political
and societal processes.
In the following, Section 2 describes the design and compilation of the project’s Danish and
German social media discourse corpus. Section 3 discusses a selection of preliminary, cor-
pus-based descriptions of linguistic and communicative characteristics of xenophobic, mainly
online hate speech in Danish and German found in the sub-projects; these ndings constitute
the basis for our further analysis. The section also presents the conceptual and methodological
bases of the approach to the perceptual analyses of hate speech. Section 4 contains the Conclu-
2. Data sources and corpus compilation
For qualitative and quantitative purposes, the project collects social media data for both Danish
and German from two of the largest players in this domain, Facebook (FB) and Twitter (TW).1
Data acquisition is carried out continuously, using the ocial APIs provided by these ser-
vices, and the harvested data is stored on secure university servers. For both social media,
data space is automatically restricted by a seeding process, but while FB uses seed pages,
TW works with individual seed word forms, and the two data sets are therefore quite dierent.
Thus, we used a list of political parties, news media and public gures for FB, while for TW
we used a combination of highly frequent words (e.g. for Danish og ‘and ’, eller ‘or’, har has’,
er ‘is’, and correspondingly the equivalent und, oder, hat, ist for German) and hate-speech
specic words (e.g. for Danish muslimer ‘muslims’, perker ‘immigrant’ (a standard derogato-
ry term), indvandrer ‘immigrant’, skide ‘shitty’, matched by the equivalent Moslems, Kanacke,
Immigrant, scheiß for German), as well as particular inected forms, such as the more frequent
plurals of person nouns. As a consequence, our FB corpus is by design more topic-restricted
than the TW corpus; the latter comes closer to a general social media corpus. In addition,
thanks to another very useful dierence, the two corpora supplement each other in many ways.
For instance, TW utterances are typically short, public and sender-driven, while FB’s posting
culture is rooted in ‘friend’ networks with a two-way communication channel. TW uses hash-
tags that can b e useful for topic-ltering, but it is tex t-only, while F B contains pictures and longer
posts, or even turn-taking comment chains, with room for argumentation and illustration.
Within the project, our corpus fulls multiple functions. First of all, it is a source of hate
speech examples for qualitative analysis; also, it is of use in the inter view and experimental sub-
tasks. Second, it helps identify slurs and linguistic patterns typical of hate speech; and third, it
allows for statistical evaluation and comparison with background data. In order to support all
these tasks and make ecient use of the corpus, the raw text data had to be ltered, linguisti-
cally processed, and turned into a searchable database with a user interface.
2.1 Preprocessing and Filtering
Apart from obvious ltering tasks, especially in the rst phase of the project (e.g., eliminating
items that were neither Danish nor German along with non-textual data or source anonymisa-
tion), we also tried to constrain content, by creating smaller sub-corpora in order to facilitate
inspection. For this purpose, we applied a boot-strapping approach with lists of key wordforms
or stems (minority groups, slurs, and “negativity” words); here, new trigger words would be
found and added to known trigger words found in previously analysed sentences. Given the low
inter-annotator agreement of human hate-speech classication (Ross et al. 2016), we are taking
care not to exclude data prematurely; the or iginal corpus is maintained for further inspection, as
well as for contextual verication of material stemming from, for instance, interviews, external
sources, and introspection. Current corpus sizes are 1,300 million and 200 million tokens for
German and Danish TW, respectively, and 200 million and 60 million for German/Danish FB.
1 Other so cial net works were also consid ered, but they turned out to be either irrelevant to the topic or to sue r
from a lack of accessib le data; alter natively, data harvesting was hamp ered by technical/le gal restri ctions.
Baumgarten et al.
2.2 Linguistic annotation
By far the most challenging task in dealing with a corpus is the linguistic annotation which is
necessary to allow corpus-linguistic methods of knowledge collection (e.g. Baker et al. 2013)
– in our case, this is the identication, quantication, and interpretation of linguistic vehicles
of hate speech. Because of the enormous size of the corpus, annotation must be performed on
a high-performance computer cluster (such as SDU’s Abacus) rather than on individual ma-
chines. We use well-established NLP tools for these tasks, the DanGram parser for Danish (visl. and GerGram for German ( Both are using the Constraint Gram-
mar (CG) formalism (Bick & Didriksen 2015), and perform lemmatization, morphological anal-
ysis, syntactic disambiguation, dependency parsing and semantic annotation of named entities
and semantic noun classes. Even though both parsers have been used in numerous corpus and
applicative tasks before, they were not built with social media data in mind; hence, features
such as incomplete sentences, spoken-language traits, orthographical errors (or creativity),
compounding and smileys/emojis pose problems for standard parsers, built for genres such as
news texts and literature. This genre challenge is one reason why using CG parsers is a good
idea: Being rule-based and lexicon-driven, they are easier to modify and adapt purposefully,
compared with statistical systems that would face a serious lack of data suitable for training,
both in Danish and German.
In the Danish example below, each word token is followed by various tag elds, covering
lemma [....], part of speech and inexion (e.g. V PR AKT for verb, present tense, active), com-
pound analysis (e.g. <N:lort~e+racist>), syntactic function (e.g. @SUBJ for subject and @ACC
for accusative object), semantic role (e.g. §AG for agent), verb frame (e.g. <fn:teach>), depend-
ency links (#n-->m) and secondary tags such as <interr> (interrogative) semantic class (e.g.
<Hideo> for “ideologicalhuman):
I [I] PERS 2P NOM @SUBJ> §COG #1->2
ved [vide] <fn:know> <mv> V PR AKT @FSSTA #2->0
intet [intet] <quant> INDP NEU S @<ACC §SOA #3->2
om [om] PRP @<PIV #4->2
$, [,] PU @PU #5->0
hvordan [hvordan] <interr> <amod> ADV @ADVL> #6->10
kvinder [kvinde] <fem> <H> N UTR P IDF NOM @SUBJ> §AG #7->10
i [i] PRP @N< #8->7
Mjølnerparken [Mjølnerparken] <top> <Lh> PROP NOM @P< §LOC #9->8
opdrager [opdrage] <fn:teach> <mv> V PR AKT @FSP< §TP #10->4
deres [de] <poss> PERS 3P GEN @>N #11->12
børn [barn] <Hbio> N NEU P IDF NOM @<ACC §PAT #12->10
$, [,] PU @PU #13->0
lorteracister [lorteracist] <N:lort~e+racist> <Hideo> N UTR P IDF NOM @VOK #14->10
[You know nothing about how women in Mjølnerparken educate their children, fucking racists]
2.3 Corpus search interface
The linguistically annotated corpus, in anonymous form and password-protected, has been
made accessible for project members at the CorpusEye site (, through a tai-
lor-made graphical user interface (GUI), internally using the Open Corpus Workbench / CQP
query language (Evert & Hardie 2011). Apart from traditional word form searches, the Cor-
pusEye GUI (Bick 2005) allows regular expressions and provides menu-driven access to
features such as lemma, syntactic function, and semantic class. For the current project, new
features were added (Bick & Didriksen 2017-), for instance, subsearches, where the main search
is performed on the output of another (ltering) search. Thus, it is possible, for instance, to
search for adjectives linked to the semantic class of “nationality nouns” in a sentence set pre-
ltered for swearwords. Results are rst shown in classical concordance format, but can then
be expanded, quantied or sorted for absolute or relative frequency of target search elds.
The example (Table 1) shows the top-ranking adjectives associated with a number of minority
nouns, as well as a list of derogative adjectives (in italics) found high on the correlate list.
[immigrant] FB: ikke-vestlig, kriminel, illegal, vestlig, muslimsk, utilpasset
TW: ikke-vestlig, illegal, arbejdsløs, vestlig, muslimsk, kriminel
tyvagtig, hjernelam, pædol, fucking, satans
[refugee] FB: såkaldt, syrisk, økonomisk, ægte, kriminel, muslimsk, palæstinensisk
TW: syrisk, palæstinensisk, sårbar, grisk, nytilkommen, såkaldt, mindreårig
[foreigner] FB: kriminel, højtuddannet, hård, uintegreret, såkaldt, herboende, ikke-mus-
TW: kriminel, højtuddannet, hjemløs, middelmådig, højtlønnet, ikke-vestlig
satans, fucking, fæl, væmmelig, forpulet, morderisk
muslim FB: rettroende, dårlig, kær, ekstremistisk, ubeviselig, ikke-vestlig
TW: religiøs, kær, moderat, sekulær, rettroende, frafalden
fucking, ulækker, sindsyg, forbandet, satans, rådden, hjernedød, bindegal
Table 1: Adjecti ve collocates of immig rant minori ty nouns.
The lists provide a rough idea of the mental space associated with the person concepts in ques-
tion: While t he rst three are all perceived as potent ially criminal, this is most promi nent for the
concept of foreigner, while immigrants are categorized on a western/non-western axis and ref-
ugees according to their legitimacy and provenance. The concept of Muslim evokes degrees of
faith, extrem ism, and the ironic kær ‘dear’. Muslim also attrac ts the largest number of defamato-
ry adjectives in t he top frequency ranks, while refugee almost goes free. Though simpli fying and
without context, even these short lists show how simple statistical corpus ndings can prompt
further qualitative research. For instance, there is a hint that foreigner is associated with danger
(kriminel ‘criminal’, hård ‘ruthless’, fæl ‘sinister’, morderisk ‘murderous’), something that con-
icts with the competing “educated resource” concept (højtuddannet ‘well-educated’, højtløn-
net ‘well-paid’), and therefore would warrant further inspection in context.
Baumgarten et al.
3. Production and perception perspectives
The XPEROHS corpus data described above is used for a number of dierent sub-projects for
both languages, investigating both the most targeted and lexically local expression of hate
speech, slurs, as well as more complex linguistic constructions and rhetorical strategies of on-
line hate speech. Finally, the corpus provides a point of departure for the interviews, question-
naires and experiments used in the empirical part of the project.
3.1 Sub-project 1: Semantics and Pragmatics of Denigration
Ethnic slurs s uch as Kike for Jews, Gypsy for Romani people, and t he infamous N-word for Blacks
are well-known terms for expressing linguistic aggression. In accordance with the narrow de-
nition in the current study, ethnic slurs are derogatory nouns intended to refer to members of
distinct ethnic groups; sadly, the use of such terms is widespread. Referring to the category as
ethnophaulisms, Rice et al. (2010) present more than two hundred English slurs targeting a va-
riety of dierent European nationalities, and examine how informants judge the degree of neg-
ativity associated with the terms.
In Danish, the two most common slurs are neger, a term for Blacks, and perker (most likely
constructed from the words perser ‘Persian’, and tyrker ‘Turk’), a derogatory term for non-Jew-
ish people of Middle Eastern or North African descent. An important dierence between these
terms is that neger has the non- derogator y counterpart sorte ‘black people’, to refer to the people
targeted by the slur, whereas perker does not seem to have such a counterpart.
Our Danish Facebook data reveal further interesting dierences. One signicant observa-
tion is that 9.7 percent of the 258 occurrences of the base lemma perker in the sub-corpus are
preceded by the adjectives fucking ‘~fucking’, skide ‘~shitty’, forpulede ‘~fucking’, ‘forbandede
‘~damned’, or the prexoid lorte- (‘~shit-’) used as a separate word, all expressing irritation, an-
ger or contempt towards the (intended) denotata of the noun modied. In contrast, only one
of the 438 occurrences of neger in the corpus is preceded by such an expression of a negative
attitude (skide). Consequently, while both terms are slurs, our corpus indicates a considerable
dierence with respect to how strongly they are associated with negative appraisals. Another
substantia l dierence relates to the distinction bet ween use and mention (see Cappelen, Lepore
& McKeever 2019). For neger and perker, this distinction corresponds to the dierence between
occurrences where the expressions are applied to people (use) versus occurrences referring to
terms applied to people (mention). Out of the total number of occurrences of perker in our cor-
pus, the proportion of mentions is much lower, 7.8 percent, compared to occurrences of neger,
20.7 percent, measured by the proportion of occurrences within quotation marks (“”), forms
of the verbs sige ‘say’ and hedde ‘be called’, or of the noun ord ‘word’. One explanation of this
dierence is that neger is frequently mentioned in excha nges where the appropriateness of using
the term is debated, and often defended. Such debates regarding the appropriateness of saying
perker are absent in our corpus. Awareness of the severely derogatory opinion communicated
by the term plausibly prevents typical language users from considering discussions of this kind
In the German data, the same tendency can be observed: Whilst Kanake, approximately cor-
responding to Dan ish perker, is hardly discussed in ter ms of appropriateness (the exception is its
use in its original meaning where Kanake designates the indigenous people of New Caledonia),
by contrast we nd many instances of meta-linguistic discussions about the meaning and use
of the word Neger in the corpora (e.g., Wort ‘word’ being the most frequent left collocate besides
the denite ar ticle, in absolute numbers FB 33, TW 65); even so, the “use”-occurrences, just like
in Danish, by far outnumber the “mention”-occurrences. Roughly the same applies to Zigeuner
‘gypsy’, cf. the following examples2 (1) and (2):
(1) De Snt und Roma heßen m Volksmund Zgeuner. Was Zieh Gauner bedeutet.
‘The Sinti and Romanis are called Zigeuner (gypsies) in common parlance. Whch
means travelng crooks.’
(2) Derzeit kursieren Zigeuner ohne Kenntnis der Ortssprache auf Bahnhöfen mit Bettel-
Zettel und hohler Hand und belästigen die Bahnfahrer.
‘At present, Zigeuner (gypsies) not knowing the local language are running around on
railway stations with a begging slip, their hands cupped, and harassing the passengers’
The meta-lingu istic discussions found in ou r corpora t within a broader ongoing debate in Ger-
man society (see e.g. Tlusty 2018) about the appropriateness of slur terms, for instance, in com-
pounds such as Negerkuss, Mohrenkopf ‘whippet cookie, lit. negro’s kiss, blackamoor’s head’ or
Zigeunerschnitzel, Zigeunersauce ‘spicy cutlet, sauce; lit. Gypsy style cutlet, sauce’, or occurring
in older children’s literature. The signicant observation is that in our FB hate speech sub-cor-
pus, we almost exclusively nd statements doubting the denigrating status of these expressions
whereas the general discussion has tended to be much more nuanced (see e.g. Neufeld 2013);
even so, the expressions in question have largely disappeared from public language use.
Regarding the German ethnophaulisms, we aim to analyse the whole inventory in terms
of meaning and use of such expressions. They can be identied with respect to the targeted
groups: thus, in addition to foreigners in a more general sense, the targets are people from ot her
European nationalities, along with Asians, Blacks, Middle Easterners etc., but also groups of
German speaking people like Austriansor East (vs. West Germans). Some ethnophaulisms are
mentioned in older descriptions3, whereas those ethnic “insulting words” (Beleidigungswörter)
identied as empirically prominent in German by Technau (2018), (e.g. Polacke ‘Pole’, Döner-
fresser ‘Turkish person; lit. eater of döner kebab’) all prove to be prominent in our data as well.
The dierent morphologica l (e.g., Nafri from Nordafrikaner ‘person f rom Northern Afr ica’) and
semantic patterns of word formation (e.g., Froschfresser ‘French person; lit. frog eater’ (cf. Frog);
Schlitzauge ‘East Asian person’; lit. slit eye’ (cf. Chink)) require a more in-depth analysis. What
can be stated already now is, however, that apart from the rather mild slur AmiUS American’,
which is the one most often occurring in both the TW and the FB corpora (> 10.000 occurrenc-
es), the most frequent and most varied ethnophaulisms are those referring to (Muslim) people
2 All examples are taken fro m the corpus a nd reprodu ced verba tim.
3 e.g. Böhmak ‘person from the Bohemia region’, Pachulke ‘Russian’ (Winkler 1994).
Baumgarten et al.
from the Maghreb and the Middle East. Amongst the semantic sources of Anti-Muslim ethno-
phaulisms are those referring to religious (e.g. Mullah, Ayatollah, Mufti) and societal (Scheich
‘sheikh’, Sultan ‘sultan’) fu nctions, in addition to common names ( Ali), stereotyp ically associat-
ed occupations (Teppichhändler ‘carpet dealer’, Dattelpücker ‘date picker’, Kameltreiber ‘camel
driver’), or ethnophaulisms referring to sex ual intercourse with ty pical animals from the region
or cultu re (Ziegen-, Schafs-, Esel-, Kamelckergoat-, sheep-, donkey-, camelfucker’). When deal-
ing with expressions like Mullah, the cha llenge is to discern whether or not for each occurrence,
in the specic context in which it occurs, the word is used as a slur or in its common denoting
function; cf. examples (3) and (4):
(3) Ein Mädchen-Killer steht auch ganz unten in der Hierarchie der Gesellschaft, aber
Merkel beharrt ja darauf die Mullahs mit ihrer Frauenfeindlichkeit massenhaft ins Land
zu lassen!
A girl killer likewise ranges at the very bottom of the societal hierarchy, but Merkel
[Angela Merkel, the German Federal Chancellor] insists on letting hordes of Mullahs
with their misogyny slip into the country.
(4) Zweitens ist auch der Kampf gegen den fundamentalistischen Fanatismus der An-
hänger des Al-Qaida-Netz werks und des Mullah Omar noch nicht gewonnen
Second, the ght against the fundamentalist fanatism of those supporting the
Al Qaeda network and Mullah Omar is not won yet.’
In addition to slurs, dehumanizing metaphors and metonyms are also part of online immigra-
tion debates (see, e.g., Böke 1997; Demjén & Hardaker 2017; Kałaszn ik 2018), which is why these
expressions are also analysed in our project.
Immigrants, refugees, Muslims and other groups are constructed in terms of dehumaniz-
ing conceptual metaphors that appeal to source domains of animals (e.g. svin / Schwein ‘pig’),
(mental) illness (e.g. syg / krank ‘ill’ or mentalt forstyrret / geistig gestört ‘mentally impaired’),
infestations (e.g. pest / Pest ‘plague’; cf. examples (5) and (6)), scum (e.g. aald / Abfall ‘waste’ or
skidt / Dreckdirt’), and natural disasters (e.g. oversvømmelse / Flut ‘ood’).
(5) Pesten Islam bringer død og ødelæggelse hvorend den får lov at orere.
The plague of Islam brings death and destruction wherever it is allowed to develop.’
(6) Wenn ch sehe, we n Deutschland sch de Islampest verbretet und wenn ch dese
Pest überall n den öentlchen Verkehrsmtteln sehe, stegt Hass und Wut n mr auf.
‘When I see how the Islamic plague is spreading in Germany and when I see this plague
everywhere in public transportation, hate and fury rise up in me.’
3.2 Sub-projects 2 and 3: The Subtle Dynamics of Hate Speech in
Danish and German
In order to allow a comparison, 2 and 3 follow the conceptual and methodological organization
of sub-project 1; as t heir only dierence is in the lang uages, they are described together. As with
sub-project 1, sub-projects 2 and 3 address all three of the project’s overall aims, but they are
concerned with a wider range of subtle mechanisms for expressing hate speech, inasmuch as
they focus on the levels of morphology, syntax, and discourse. The specic aims of the sub-pro-
jects are
a) to identify, based on corpus and interpretative analyses together w ith user percep-
tions, a core (prototypical) repertoire of interactional meanings and patterns for
hate speech in the two languages;
b) to disting uish between this ‘indisputable’ hate speech and interactional meanings
and patterns that can be considered as less clear cases of hate speech – or even,
depending on the context, as not being hate speech at all (Meibauer 2012).
In particular, sub-project 2 represents the rst in-depth pragmatic investigation of this kind of
hate speech for Danish, while sub-project 3 builds on existing work in German but develops this
further, moving beyond explicit anti-Semitic and ethnic pejorative expressions.
Both in Danish and German, we observe parallel means of expression at various linguistic
levels. These recurring expressions can be conceived of as constructions (see Geyer 2018). An
example with a high recognition factor is the I am no racist but-construction (Jeg er ikke rac-
ist, men … / Ich bin kein Rassist, aber … )4 that combines two statements expressing adversative
meanings, where the rst statement serves to signal the speaker’s pretended reecting mind,
and thus hedges the second statement (whose completion may turn out to be qu ite oensive). In
terms of quantity, the number of occurrences of this pragmatic construction is comparable in
the two sub-corpora: 190 examples were found in Danish (FB: 81 TW: 109) and 340 in German
(FB: 84, TW: 256). Some examples follow:
(7) Jeg er ikke racist, men realist. Hvad gør Danmark for de ældre på plejehjem, som igen-
nem tiderne har slidt sig selv op ved hårdt arbejde og har opbygget vort velfærdsamfund?
Hvad har de såkaldte ygtninge bidraget med til samfundet. Kriminalitet i rå mængder.
‘I’m not a racist, but a realist. What does Denmark do for the elderly in nursing homes,
who over the years have worn themselves out by hard work and have built up our welfare
society? What have the so -called refugees contributed to soc iety? Cri me in raw quantities.’
4 Other expressions llin g the slot oc cupied by racist in the canonical construction are fremmedhader ‘xenopho-
be’ or radikal ‘radical’ in Danish, and Antisemitanti-Semite’ Hater ‘hateful person’, or Rechter ‘right-winger’ in
Baumgarten et al.
(8) Ich bin kein Rassist aber ich bin strikt dagegen das Ausländer im Ausland in die Politik
aufgenommen werden. Egal welches Land das bringt nur Unheil mit sich.
‘I am not a racist but I am strictly against foreigners abroad being allowed to engage in
local politics. No matter in which country, this only brings disaster.’
Another, simila r construction which seems to be a common strategy for disseminating negative
stereotypes is the phrase Jeg har ikke noget imod …, men / Ich habe nichts gegen …, aber (‘I have
nothing against…, but’). In Danish, there are not many examples with regard to Muslim (FB:
4; TW: 2) or jøde ‘Jew’ (FB: 2; TW: 0). By contrast, this construction is found far more often in
(9) Jeg har ikke noget imod muslimer, det har jeg vitterlig ikke. Men jeg er virkelig træt af at
de skal særbehandles ud fra deres religion. Hvis vi nu vendte fortegnet, ville vi så også få
særbehandling i deres hjemland?? NEJ! Så hvorfor er det så, så svært at forstå? De lever
og ånder for deres religion. Vi lever og ånder for vores principper, værdier, lovgivning,
traditioner, og ikke mindst det åbne samfund, hvor vi respektere hinanden på en
præsentabel måde.
‘I don’t mind Muslims, really I don’t. But I am really tired of their being preferentially
treated because of their religion. Putting it the other way ‘round, would we too
get preferential treatment in their home country?? NO! So why is this so, so hard
to understand? They live and breathe for their religion. We live and breathe for
our principles, values, laws, traditions, and not least an open society, where
we respect each other in a respectable way.’
(10) Was muss eigentlich noch passieren das Deutschland aufwacht??!!! Bald liegt hier alles
in Schutt und Asche... das sollten wir uns mal in anderen Ländern erlauben da wirst er-
schossen... Ich habe nix gegen Ausländer aber dieses Pack kann wegen mir wieder dahin
wo se her komm...
‘Really, what more has to happen to awaken Germany??!!! Soon everything here will be
in ruins and ashes... imagine we would allow ourselves to behave like this in other coun-
tries, you would get shot … I have nothing against foreigners but, as for me, this pack can
go back to where they came from...’
In German, the most frequently mentioned target group in this construction is Ausländer ‘for-
eigners’ (FB: 65; TW: 42), followed by Juden ‘Jews’ (FB: 33; TW: 24) and Flüchtlinge ‘refugees’
(FB: 27; TW: 11). Examples with Muslime ‘Muslims’, as in (11), are also comparatively rare (FB:
7; TW: 8).
(11) Ich habe nichts gegen Moslems, so lange sie nicht in Deutschland leben. Überall in der Welt,
sorgen sie ür Unfrieden.
‘I have nothing against Muslims, as long as they do not live in Germany. Everywhere in
the world, they cause discord.’
In German, this construction is predominantly used to express negative attitudes, especially
against people with a dierent ethnic origin, against dissenters, and against homosexuals.
Another research objective in sub-projects 2 and 3 is to study the use of irony. Here, we anal-
yse the dierent ways the opposite of a literal expression is conveyed. An example of irony is the
construction die ach so / de åh (‘the oh so’) + (often positive) adjective + noun. While it is used
quite often in Ger man, it is rarely found in the Danish cor pus (12 results in total). In addit ion, the
use of this combination (adjective + noun in the oh so -construction) is quite specic in German:
the noun most commonly used is Flüchtlinge ‘refugees’, primarily combined with the adjective
arm ‘poor’ (FB: 16 occurrences; TW: 14 occurrences), as in example (12):
(12) Diese ewige Di skussion hier über di e ach so armen Flüchtlinge die eigentlich keine Flüchtlinge
sind. Wenn illegale das Land zu verlassen haben, dann haben die der Auorderung Folge zu
leisten. Punkt aus... dann soll die Polizei ihre Dienstwae benutzen… So einfach ist das.
This never-ending discussion here about the oh so poor refugees who are not really ref-
ugees. If illegal [immig rants] have to leave the countr y, then they must follow that order,
period [and] then the police should use their service weapon ... It’s that simple.’
Another relatively frequent occurrence within the ach so -construction is the combination of the
adjective friedlich ‘peaceful’ together with the nouns Islam (FB: 7 hits; TW: 20 hits), Muslime, or
Moslems (‘Muslims’), as in der ach so f riedliche Islam or die ach so friedlichen Muslim e ‘the ever so
peaceful Islam/Muslims’ (FB: 2 hits; TW: 6 hits). The speaker’s intention is to create a general
association between Muslims and the Islamic religion on the one hand, and terror and violence
on the other.
In addition to the hate speech examples, one also nds the oh so-construction used in ‘coun-
ter speech’ (i.e., to counter a claim), as in (13) and (14):
(13) Grænsekontrollen koster 250 millioner kroner om året og har kun reduceret tallet af
asylansøgere en 1/4 del. Tror du kun det er de åh så skrækkelige udlændinge der skal
igennem den? Nej, det er også danskere som mig der bor i udlandet. Super brug af de
mange penge!
‘The border control costs DKK 250 million a year and has only reduced the number of
asylum seekers by one-fourth. Do you think it’s only the oh so horrible foreigners who
have to pass it? No, it is also Danes like me who live abroad. A terric use of those huge
Baumgarten et al.
(14) Ihr regt euch über die ach so bösen Moslems auf denen ihr immer unterstellt in die
Opferrolle zu gehen dabei seid ihr es doch
‘You are nervous about the oh-so-evil Muslims and take it that you always have to as-
sume the role of the victim – which of course in fact you are’
A related construction (FB: 67; TW: 107) is die so genannten Flüchtlinge ‘the so-called refugees’
which occurs frequently in German. It occurs also in Danish de såkaldte ygtninge on Facebook
and Twitter, but not as often as it does in German (FB: 6; TW: 4); the use of so-called downplays
the status of the refugees as a group.
We also nd expressions in both languages stating that foreign groups should leave the
country. In Danish as well as in German, a very common collocate of ud / raus ‘out’ is the word
udlændinge / Ausländer ‘foreigners’. Thus, in our corpus, the combination udlændinge ud ‘out
with the foreigners’ occurs 53 times in the Twitter section and 22 times in the Facebook part.
The German Ausländer raus is also very frequent (FB: 212; T W: 305; cf. (15)), though considering
the larger German corpus, not signicantly so. The related ygtnin gene u d / Flüchtlinge raus ‘out
with the refugees’ (overall results in Danish: 49; in German: 185) or muslimerne ud / Muslime
raus ‘out with the Muslims’ (overall results in Danish: 85; in German: 108) can also often be
found in both languages. Common usage also includes constructions with verbs such as smide
‘throw’ or sende ‘send’, especially for Danish like in (16):
(15) Alle Ausländer raus hier geht euer land auauen Ihr kommt doch nur her weil es geld
gibt und bitte nimmt eure kopf tuch mädels mit
‘All foreigners get out of here. Go and build up your cou ntr y. You only come here because
there is money. And please take your headscarf girls with you’
(16) De høre ikke til her, vi vil ikke have blandet hverken blod eller religion. Og vi vil slet ikke
have deres krig. Smid dem ud hurtigst mulig inden de overtager vores land
‘They do not belong here, we do not want to blend blood or religion. And we den itely do
not want their war. Throw them out as fast as possible before they take over our country’
While Ausländer raus has established itself as a slogan or catchphrase, there is another syntac-
tic construction, in German as well as in Danish, in which the adverb raus is combined with a
prepositional phrase identifying the target group, as in Raus mit den Flüchtlingen ‘Out with the
refugees’. However, such constructions (with target groups like Flüchtlinge (22), Ausländer (12),
Migranten (11), Asylanten (8) etc.), are poorly evidenced in our corpus. In this context, an often
used noun is the pejorative Pack ‘vermin’ (FB: 268; TW: 283); compare also other compound
nouns with Pack, in particular Dreckspack ‘pesky varmint’ (FB: 54; TW: 25).
In Danish, t he noun pak itself as well as compounds with this word (as in rakkerpak ‘outcast’)
are less common. We obtained 8 resu lts for expressions like Ud med det pak! and 12 resu lts for Ud
med det rakkerpak! The most frequent combination in Danish is Ud med det lort! ‘Out with that
shit’ (FB: 90; TW: 4). Unexpectedly, for German only 4 results can be found for the analogue
example Raus mit dem Scheiß!
In summary, it can be stated that in German and in Danish, similar incentives can be found,
irrespective of their syntactic construction. Only in particular contexts, the group which is sup-
posed to leave the country is specied. In many of the examples, the concrete appeal only con-
tains the abusive terms (e.g., ‘shit’ or ‘vermin’), and not the target group itself; the equating of
‘shit’ with, for instance, Muslims is left to the context; the respective group is considered to be
inferior and not worthy of staying, and is treated as such.
3.3 The Ph.D.-project: Rhetorical Strategies in Danish Online Hate
The Ph.D.-project focuses on the dynamics of hate speech and strategies for constructing evi-
dentiality. One crucial research question is whether the presentation of certain topics seems to
initiate hateful comments. A report from PET’s Center for Terroranalyse (2008: 5) concludes
that stereotyping of minority groups and emotional metaphors (e.g., ‘holy warrior’, ‘martyr’)
in online contexts seem to provoke hateful speech that might lead to real-world violent hate
crimes. In 2017, the Danish Institut for Menneskerettigheder (‘Institute for Human Rights’) de-
rived equivalent conclusions in a repor t on the initiation and dynamics of hate speech observed
on the Twitter and Facebook pages of the Danish media channels DR TV and T V2 News (Zuleta
& Burkal 2017). The institute registered that especially topics on religion, faith, refugees, equal-
ity, politics, and integration triggered hostile rhetoric. The Ph.D.-project elaborates on these
observations, but incorporates more specic linguistic perspectives in order to gain a deeper
insight into the dynamics of the recontextualization processes of hate speech, especially with
regard to the commentators’ use of evidentiality. (In this connection, evidential strategies are
dened as the commentators’ ‘I have heard’, ‘I saw’, and other such expressions as legitimation
of their statements; Mushin 2013).
In online hate speech, hyperlinks are often used as an evidentiality tool. By using this strat-
egy, the author removes the focus from the utterance to the content of the link, which leads to
a complication of the communicational context. Furthermore, the hyperlinks referred to are
often either ‘blind’, or the reader may not be able to activate or check them, such that a false
‘documentation’ can occur.
The overall organizational patterns of online communication are of certain interest for sev-
eral reasons (here, the project also focuses on the role of ‘counter speech’, see above). First and
foremost, these pattern s are determined by the aordances li nked to the medium (Jensen 201 4),
but – in addition to the technical restrictions and facilities – it is hypothesized that especially
sensitive topics (e.g., religion, ethnicity) have a co-determining inuence on the organization
of the dialogue and the way the participants position themselves and are positioned grammat-
ically, semantically, and pragmatically. Thus, recent research has already pointed out that the
cooperative maxims are outed in online hate speech communication (Jensen 2014).
Baumgarten et al.
The dynamics of hostile rhetoric and the complexity of the online dialogues are exemplied
in the example below. The comments were posted in relation to a documentary about people
smuggling on DR’s TV channel (the parentheses in the left-hand column indicate speaker ini-
(TS) vi skal ikke have Isis eller andre kriminelle ind i Europa. De er økonomiske migranter
der tager vores penge. Dem der vil hjælpe dem kan tage ned og hjælpe dem i deres land.
Europa er ikke et toilet som nege r og muslimer bare kan komme og skide i. Og det er prob-
lemet med de este indvandrer, De har ingen respekt for at Europa er for europæer. (…)
Ham der + krimin elle indvandrer skal smides u d af Europa. _ Hvorfor tror I det k un er de
hvides lande der skal være multikulturelle ? Do the research and you shall nd .. https://
‘we don’t want to let ISIS or other criminals enter Europe. They are economic mi-
grants who take our money. Those who want to help them can go and help them in
their own countries. Europe is not a toilet that negroes and Muslims can just come
and shit in. And that is the problem with most immigrants, they do not respect that
Europe is for Europeans (…). That guy + criminal immigrants should be thrown out
of Europe. Why do you think that it is only the white people’s countries that should
be multicultural? Do the research and you shall nd ...
(IA) Det be viser igen at de ( skider ) på reglerne ,
‘That shows again that they shit on the rules’
(KID) Hvem er “ de”?
det virker som om du er en, der generaliserer...
‘Who are “they”?
. It seems like you are somebody who generalizes …’
(IA) nej ,talemåde
‘no, a way of speaking
(SJP) og det gør de jo også he r i DK de har d eres egne regler
ok bastarder he le bundet
‘and they do that a s well here in DK they have thei r own rules
pack of ba stards the whole lot
In the extract, the organization of the dialogue and the lack of cooperation makes it possible
for a participant to ignore a withdrawal (nej, talemåde ‘no, way of speaking’) and resume the
hostile rhetoric. In the example above, the ambiguity of hostile content uttered by (TS) is con-
textualized by (IA) in terms that might lead to escalation (de ‘they’) but is contested by (KID)’s
objection criticising the generalization. (IA) then cooperates and defuses his/her contribution.
The next turn then, however, uttered by (SJP), outs the Cooperative Principle, as the commen-
tator both overrides the withdrawal and defuses the hostile rhetoric – but in an escalating way,
so that the road is open again for hateful comments. At the same time, the extract exemplies
the dic ulties encountered when analysing onli ne comments as if they were dialogues: we can-
not be sure if (IA)s comment addresses the content of the hyperlink, but since he/she does not
defend the generalizing comment, we deem it probable that ‘they’ relates to the comment itself.
In the above example, a link to a video on You Tube is used as an evidential strategy. The video
will – according to TS – ‘prove’ that foreigners do not respect European countries, and this is the
reason why only these countries are becoming multicultural.
The PhD project in question is based on the theory of integrative pragmatics (Culpeper and
Haugh 2014) but incorporates elements from Critical Discourse Analysis (CDA; Fairclough
1992, Wodak 2007, Wodak & Reisigl 2015). The CDA methods have been taken aboard in order
to turn t he attention to the relationship between discou rsal and social changes and t he potential
consequences of hate speech for individuals and groups. Social isolation and loss of dignity in
society are considered (Nilsen 2014: 8) as well as t he potential connection to hate cri me and ter-
ror (Center for Terroranalyse / PET 2008). Finally, further issues to be addressed in the project
involve censorship, freedom of speech, and the democratization of discursive rights.
3.4 Sub-project 4: Instrumental approaches to perceptions of spoken
and written modes of hate speech
Hate speech is not exclusively a matter of written language, although much of the current re-
search is focused on written hate speech – particularly so in the social media discourse. At the
same time, much of the current research focuses on the production of hate speech, although
its reception by readers and listeners is arguably just as important. Contrastive and perceptual
analyses of spoken and written hate speech are, therefore, necessary to provide a more accu-
rate and comprehensive description of the nature of the phenomenon: for instance, what people
react to specically when they read or hear hateful messages, where they place the boundaries
between hate speech and ‘acceptable’ forms of negative expressions, and whether or not the
written mode (i.e., reading) creates a personal detachment from perceptions of the hatefulness
of the content that does not exist in the same way in the spoken mode (i.e., listening). Gain-
ing these insights will enable us to describe and theorize the interdependence of the linguistic,
communicative, and perceptual dimensions of hate speech.
Two main questions are addressed in the sub-project: First, is the perception of hate speech
similar across written and spoken language? And secondly, is it primarily the words that deter-
mine the perception of hate speech or does prosody (i.e., speech melody and voice quality) play
a role as well, e.g. to the extent that w ritten hate speech becomes acceptable in spoken langu age,
or conversely, that acceptable written language becomes hate speech in spoken language? We
approach these questions through an innovative multiple methods design that combines im-
plicit and explicit instrumental measurements as well as quantitative and qualitative analyses.
It is known that prosody (often called people’s oldest means of acoustic communication, cf.
Gussenhoven 2004) is directly linked to listeners’ interpretations of speaker traits, attitudes,
and emotions (Bänzinger & Scherer 2005; Da Silva & Barbosa 2017; Niebuhr 2017; Neitsch
2019). As Cabane (2012: 136) puts it: “Speech melody is hardwired in our brains.” If meanings
Baumgarten et al.
conveyed by prosody contradict those conveyed by words, listeners give melodic meanings pri-
ority over lexical ones and interpret the corresponding verbal utterances as non-sincere, i.e.,
ironic or sarcastic (Landgraf 2014). Hence, it is reasonable to assume that, depending on how
prosody interacts with the coinciding words (supports or undermines them), we will nd signif-
icant dierences in what is perceived as hate speech in written and spoken language. In spoken
langua ge, it seems possible to manipulate prosody to downplay writ ten hate speech to the extent
that it is not even rated as hate speech anymore; even so, it is likely that there is a limit for this
manipulat ion to be possible. The limit may be determined by the sema ntic content and the emo-
tional load of particular key words expressing hate (epithets, swearwords etc.), by the societal
sensitivity of the topic that is referred to and/or by the recipient, their age, gender, personality
(i.e., the “big 5”; John et al. 2008), language background, and social status.
The two questions raised above are addressed based on empirically-derived stimuli of writ-
ten and spoken hate-speech tokens from the other project modules. The stimuli set includes
lexical, grammatical, semantic, propositional and rhetorical variants which are specic to par-
ticular target groups, variants that occur with dierent groups, and variants on a scale from
ambiguous to extreme. We start from a broad set of authentic written stimuli (approx. 150 to-
kens, max. 170 characters long) that undergo iterative testing for perceptual eects (Figure 1).
Figure 1. A test sequence.
Stimuli are used in both their original and manipulated forms (e.g., exchanging key words or
changing local and global prosodic characteristics towards and away from hate speech). The
spoken stimuli are produced by trained actors. Prosody manipulation is done by PSOLA resyn-
thesis (Mouli nes & Charpentier 1990) on the basis of existi ng knowledge about the phonetics of
negative emotions and expressive lexical stress, impoliteness, dominance, and irony (Poggi &
D’Errico 2018; Niebuhr 2010; Neitsch 2019).
The ‘online heat map test’ is an explorative pretest a nalysing a large stimul i set for perceptu al
eects on the basis of a heat map (Figure 2). From this test, a smaller stimuli set for tests with
combinations of physiological and cognitive measurement is derived. The latter experiment
stage consists of two tests (EX-1, EX-2), each followed by a stimulated recall interview. EX-1
presents stimuli in modality A (written), whereas EX-2 presents equivalent stimuli in modality
B (spoken). In the interviews, participants are asked to recall and comment on selected ratings
and the rating process. This provides a third dimension of hate speech perception through re-
ective accounts of exposure to hate speech, which may or may not coincide with the experi-
mentally elicited cognitive and physiological responses.
Figure 2: Two-dimensional heat map used for explicit ratings.
Four dierent test variants are conducted, one with written and spoken language, one with the
inverse order of modalities, and two further variants in which the two orders of modalities are
cross-combined with two dierent types of tests. One test measures the perceivers’ implicit re-
actions to the stimuli, whereas the other takes explicit measurements based on a rating task
Baumgarten et al.
in a 2D heat map setting (Figure 2). By clicking on the heat map, the perceived degree of hate
speech is measured in terms of the combined attributes “dislike” (x-axis: eliciting judgments of
individual a ect) and “unacceptable/not-licenced” (y-ax is: eliciting judgments of how tolerable
a stimulus is with reference to the perceivers’ understanding of operative societal norms, con-
ventions and values; Martin & White 2005). Perceivers will represent a cross-section of society
in terms of age, gender, and education.
By combining implicit physiological measurements (i.e., heart rate (HR), breathing patterns
and amplitudes using Respiratory Inductance Plethysmography (RIP)) with explicit ratings, we
cover two dierent reaction types to hate speech that people are confronted with in everyday
life. More specica lly, we investigate spontaneous and evolutionary ‘ hardwired’ reactions af ter
(incidentally) observing or reading hate speech, as well as conscious reections and judgments
on hate speech that involve given word labels and social and cultural conventions. Hence, our
experiments are the rst to determine if and how the mere action of explicitly dealing with hate
speech already chan ges people’s perception of, and reaction to it, and whether t he order in which
the two reactions are elicited – rst implicit, then explicit and vice versa – matters as well. Both
are important aspects that can help explain why established instruments of the social sciences
(e.g., surveys, interviews) increasingly and often fail to predict people’s opinions, attitudes, and
behaviour (compare the Brexit vote, or the latest (2016) US presidential elections).
EX-2 specically investigates the perception of a subset of clear and borderline hate speech
stimuli by using pupillometry based on eye-tracking. Unlike RIP/HR, pupillometry shows in
more detail for which words within the written stimuli hate-speech reactions were triggered
and gives, independently of the stimulus modality, better temporal resolution of participants’
hate-speech reactions.
4. Conclusion
The XPEROHS-project oers a comprehensive, empirically grounded, multi-method and bi-
lingual approach to hate speech in online media. In this article, we have described how a text
corpus built from Facebook and Twitter data can be used to discover linguistic patterns and
expressions of hate speech. Thus, we have presented a list of typical stereotypes and metaphors
contributing to slur words targeting minority groups (e.g., dehumanization, illness, stupidity,
pest and other animals). In addition, more subtle and indirect mechanisms, working above and
beyond the word level, were also investigated. In the experimental section, we discussed how
graded example stimuli for interviews and questionnaires are used to examine the perception
side of hate speech by ordinary language users. We argue that various non-literal and non-ver-
bal factors, such as modality and prosody, have an inuence on the perception of hate speech,
and can be captured objectively using heat maps and physiological measurements.
Both types of data – corpus-linguistic and experimental – are evaluated quantitatively and
qualitatively. For instance, we are identifying not only the range of demeaning attributes, slurs,
and othering mechanism used in minority-targeting discourse, but also their relative distribu-
tion against each other and background data. While much of the data is stored and explored
with only a linguistic context in mind, one sub-project, in particular, transcends this scope, ex-
amining entire comment threads in their original setting, including pictures, layout, and coun-
ter speech, trying to identify rhetorical strategies in online hate speech discourse.
It is an important aspect of the entire project that it is systematically bilingual for Danish
and German, allowing us to directly compare and contrast the mechanisms, lexical spread and
severity of online hate speech in these linguistically and cu lturally closely related languages. In
our contribution, we have identied such parallels for areas such as the use of irony, adversative
expressions, derogatory expressions and “leave the countr y” imperatives.
During the project, we hope to lay a foundation not only for a better linguistic and commu-
nicative understanding of online hate speech, but also for a more informed treatment of its var-
ious manifestations in terms of policy, societal harms, and pedagogics.
Assimakopoulos, Stavros, Fabienne H. Baider & Sharon Millar. 2017. Online Hate Speech in the
European Union: A Discourse-Analytic Perspective. Cham: Springer.
Baker, Paul, Costas Gabrielatos & Tony McEnery. 2013. Sketching Muslims: A Corpus Driven
Analysis of Representations around the Word “Muslim” in the British Press 1998-2009.
Applied Linguistics, 34(3). 255-278.
Bänziger, Tanja & Klaus Scherer. 2005. The Role of Intonation in Emotional Expressions. Speech
Communication 46. 252-267.
Bick, Eckhard. 2005. CorpusEye: Et brugervenligt web-interface for grammatisk opmærkede
korpora, In Peter Widell & Mette Kunøe (eds.) Proceedings 10. Møde om Udforskningen af
Dansk Sprog 7.-8. okt. 2004. Århus University. 46-57.
Bick, Eckhard & Tino Didriksen. 2015. CG-3 – Beyond Classical Constraint Grammar. In Beáta
Megyesi, (ed.) Proceedings of NODALIDA 2015, May 11-13, 2015, Vilnius, Lithuania. Linköping:
LiU Electronic Press. 31-39.
Bick, Eckhard & Tino Didriksen. 2017-. VISL CorpusEye. ><.
Böke, Karn. 1997. De “Invason“ aus den “Armenhäusern Europas“. Metaphern m Enwan-
derungsdskurs. In Matthas Jung, Martn Wengeler & Karn Böke (eds.) De Sprache des
Mgratonsdskurses. Das Reden über “Ausländer“ n Meden, Poltk und Alltag. Opladen: West-
deutscher Verlag. 164-193.
Brown, Alexander. 2017. What is hate speech? Part 1: The Myth of Hate. Law and Philosophy 36.
Cabane, Olivia Fox. 2012. The Charisma Myth: How Anyone Can Master the Art and Science of
Personal Magnetism. New York: Penguin.
Cappelen, Herman, Ernest Lepore & Matthew McKeever. 2019. “Quotation”. In Edward N.
Zalta (ed.) The Stanford Encyclopedia of Philosophy (Spring 2019 Edition), <https://plato.>.
Center for Terroranalyse/PET. 2008. Sprogbrug og terrorisme. http://www.dr .dk/NR/rdon-
Culpeper, Jonathan & Michael Haugh. 2014: Pragmatics and the English Language. London: Pal-
grave Macmillan.
Baumgarten et al.
da Silva, Wellington & Plinio Almeida Barbosa. 2017. Perception of emotional prosody: investi-
gating the relation between the discrete and dimensional approaches to emotions. Revista de
Estudios da Linguagem 25(3). 1075-1103.
Demjén, Zsóa & Claire Hardaker. 2017. Metaphor, impoliteness, and oence in online com-
munication. In Elena Semino & Zsóa Demjén (eds.) The Routledge Handbook of Metaphor
and Language. London: Routledge. 353-367.
COE = Council of Europe, European Commission against Racism and Intolerance. 2015.
General Policy Recommendation No. 15.
Evert, Stefan & Andrew Hardie. 2011. Twenty-rst century Corpus Workbench: Updating a
query architecture for the new millennium. In Proceedings of the Corpus Linguistics 2011 con-
ference, University of Birmingham, UK.
Fairclough, Norman. 1992. Discourse and Social Change. Cambridge: Polity Press.
Foxman, Abe & Christopher Wolf. 2013. Viral Hate: Containing Its Spread on the Internet. New
York: St. Martin’s Press.
Geyer, Klaus. 2018. Hadetalens ’grammatik’. Ny Forskning i Grammatik 25. 16-33.
Gussenhoven, Carlos. 2004. The Phonolog of Tone and Intonation. Sound les. Cambridge:
Cambridge University Press.
Haas, John. 2012. Hate Speech and Stereotypic Talk. In Howard Giles, Cynthia Gallois, Jake
Harwood, Miles Hewstone, Michael Hogg, Scott A. Reid & John C. Turner (eds.) The Hand-
book of Intergroup Communication. London: Routledge. 128-140.
Herz, Michael & Peter Molnar (eds.). 2012. The Content and Context of Hate Speech. Rethinking
Regulation and Responses. Cambridge: Cambridge University Press.
Jensen, Eva Skafte. 2014. Tale er tale; skrift er skrift. Om skriftsproget i de nye medier. Nydanske
Sprogstudier 46. 11-38.
John, Oliver P., Laura P. Naumann & Christopher J. Soto. 2008. Paradigm Shift to the Integra-
tive Big Five Trait Taxonomy. Handbook of Personality Theory and Research. In Oliver P.
John, Richard W. Robins, & Lawrence A. Pervin (eds.) Handbook of personality: Theory and
research. New York: Guilford Press. 114-158.
Kałasznik, Marcelina. 2018. Pejorative Metaphern im Flüchtlingsdiskurs. In Fabian Szczęk,
Joachim Klinker & Joanna Scharloth (eds.) Sprachlche Gewalt. Formen und Eekte von Peor-
serung, verbaler Aggresson und Hassrede. Stuttgart: Metzler. 67-80.
Landgraf, Rabea. 2014. Are you serious? Irony and the perception of emphatic intensication.
Proc. 4th International Symposium on Tonal Aspects of Languages (TAL). Nijmegen, Neth-
erlands. 91-94.
Martin, J. R. & White, Peter R.R. 2005. The Language of Evaluation: Appraisal in English. London:
Meibauer, Jörg. 2012. What is a context? Theoretical and empirical evidence. In Rita Finkbein-
er, Jörg Meibauer & Petra B. Schumacher (eds.) What is a context? Linguistic approaches and
challenges. Amsterdam: Benjamins. 9-32
Moulines, Eric & Francis Charpentier. 1990. Pitch-synchronous waveform processing tech-
niques for text-to-speech synthesis using diphones. Speech Communication 9. 453-467.
Mushin, Ilana. 2013. “Making knowledge visible in discourse: Implications for the study of lin-
guistic evidentiality”. Discourse Studies 15( 5). 627-645.
Neitsch, Jana. 2019. Who cares about context and attitude? Prosodic variation in the production
and perception of rhetorical questions in German (Unpublished doctoral dissertation), Uni-
versity of Konstanz, Department of Linguistics, Konstanz, Germany.
Neufeld, Dialika. 2013. It’s Time to Remove Racism from Children’s Books. Spegel Onlne Inter-
natonal, https://www.spnternatonal/germany/why-racsm-should-be-removed-
Niebuhr, Oliver. 2010. On the phonetics of intensifying emphasis in German. Phonetica 67. 170-
Niebuhr, Oliver. 2017. Clear Speech – Mere Speech? How segmental and prosodic speech reduc-
tion shape the impression that speakers create on listeners. Proceedings of the 18th Interna-
tional Interspeech Conference, Stockholm, Sweden. 894-898.
Nilsen, Anne Birgitta. 2014. Hatprat. Oslo: Cappelen Dam.
Perry, Barbara. 2005. A crime by another name: the semantics of hate. Journal of Hate Studies
4(1). 121-137.
Poggi, Isabella & Francesca D’Errico. 2018. Feeling Oended: A Blow to Our Image and Our
Social Relationships. Frontiers in Psycholog 8. 2221. doi:10.3389/fpsyg.2017.02221.
Rice, Diane. R., Dominic Abrams, Constantina Badea, Gerd Bohner, Andrea Carnaghi, Lyud-
mila. I. Dementi, Kevin Durkin, Bea Ehmann, Gordon Hodson, Dogan Kokdemir , Jaume
Masip, Aidan Moran, Margit E. Oswald, Jaap W. Ouwerker, Rolf Reber, Jonathan Schroeder,
Katerina Tasiopoulou & Jerzy Trzebinski. 2010. What Did You Just Call Me? European and
American Ratings of the Valence of Ethnophaulisms. Journal of Language and Social Psycho-
log 29(1). 117-131.
Ross, Björn, Michael Rist, Guillermo Carbonell, Benjamin Cabrera, Nils Kurowsky & Michael
Wojatzki. 2016. Measuring the Reliability of Hate Speech Annotations: The Case of the Eu-
ropean Refugee Crisis. Proceedings of NLP4CMC III: 3rd Workshop on Natural Language Pro-
cessing for Computer-Mediated Communication (Bochum). Bochumer Lngustsche Arbets-
berchte, 17, Sept. 2016. 6-9.
Technau, Börn. 2018. Beledgungswörter: De Semantk und Pragmatk peoratver Personenbe-
zechnungen. Berln: de Gruyter.
Tlusty, Ann-Krstn. 2018. Sprache st nur ene von velen Baustellen. In Zet Onlne, https://
Wnkler, Andreas. 1994. Ethnsche Schmpfwörter und übertragener Gebrauch von Ethnka.
Muttersprache 104(4). 320–337.
Wodak, Ruth. 2007. Pragmatics and Critical Discourse Analysis. A cross-disciplinary inquiry.
Pragmatics & Cognition 15(1). https://www.research-
ciplinary_inquiry. 203-225.
Baumgarten et al.
Wodak, Ruth & Martin Reisigl. 2015. Discourse and racism. In Deborah Tannen, Heidi Hamil-
ton & Deborah Schirin (eds.) The Handbook of Discourse Analysis (2nd ed.). Chichester: John
Wiley & Sons. 576-596.
Zuleta, Lumi & Rasmus Burkal. 2017. Hadefulde ytringer i den oentlige online debat. Copenha-
gen: Institut for Menneskerettigheder.
... B. als Ironie, figurative Sprache oder Imperativ (z. B. Baumgarten et al., 2019;Bick, 2020;Mondal et al., 2017), und sie tritt zudem mit sehr unterschiedlicher Deutlichkeit auf, was eine Identifizierung zusätzlich erschwert. ...
... Vor diesem Hintergrund umfasste unser von der VELUX-Stiftung finanziertes Forschungsprojekt XPEROHS (Towards Balance and Boundaries in Public Discourse: Expressing and Perceiving Online Hate Speech) auch ein Modul über die vergleichende Analyse geschriebener und gesprochener Hassrede im Deutschen und Dänischen (Baumgarten et al., 2019;Bick & Didriksen, 2015 Zuletzt zeigt diese Studie auch, dass Stimuli, die sich gegen die Zielgruppe der Muslime richteten, in beiden Dimensionen signifikant höher bewertet wurden, als wenn die Hassrede auf die Zielgruppe der Ausländer im Allgemeinen ausgerichtet war. ...
Full-text available
Zusammenfassung Dass Hassrede (hate speech) zunehmend als Problem gilt, geht nicht allein auf ein steigendes Vorkommen zurück, sondern auch auf eine erhöhte Sensibilität für dieses Thema. Da die sprachliche Struktur von Hassrede sehr vielfältig und ihre Wahrnehmung sehr komplex ist, rückt ihre Erforschung zudem verstärkt in den Fokus der Linguistik und der Prosodieforschung. In unserem Beitrag fassen wir die Untersuchung unterschiedlicher geschriebener und gesprochener Hassredetypen im Deutschen über die letzten drei Jahre zusammen. Wir zeigen, dass geschriebene Hassrede anders wahrgenommen wird, sobald man sie laut ausspricht. Dabei werden lexikalisch vormarkierte Typen von Hassrede, etwa Imperative oder solche mit Holocaust-Bezug, in ihrer negativen Wirkung verstärkt, während Hassrede, die auf stimmlichen Faktoren basiert, wie Ironie oder rhetorische Fragen, an negativer Wirkung verliert. Wir zeigen außerdem, wie sich diese Urteile in menschlichen Biosignalen wiederfinden, z. B. in EEG-Messungen zu Stress und Emotionen im präfrontalen Kortex. In diesem Zusammenhang beschreibt der Beitrag auch ein neues EEG-Experiment, das die Rolle des sozialen Kontextes auf die Wirkung von Hassrede untersucht. Unsere Ergebnisse zeigen ein höheres EEG-Stresslevel, wenn Rezipienten alleine mit Hassrede konfrontiert sind im Vergleich zur Hassrede-Konfrontation in Gesellschaft eines bekannten Menschen. Abschließend leitet der Beitrag aus allen Ergebnissen Ansatzpunkte für den praktischen Umgang mit Hassrede und deren weitere Erforschung ab.
... Rein quantitativ arbeiten dagegen Baumgarten et al. (2019) und ElSherief et al. (2018. Konkrete Methoden, die bei quantitativer Hate-Speech-Forschung zum Tragen kommen, sind beispielsweise Keywordanalyse (z. ...
... B. ElSherief et al., 2018). Abwertende ethnische Bezeichnungen (Ethnophaulismen) spielen vor allem inBaumgarten et al. (2019) eine prominente Rolle. Sie umfassen zum Beispiel deutsche Bezeichnungen wie Kanake, Nafri, Polacke oder Zigeuner oder auch gegen Deutsche gerichtete Bezeichnungen wie Alman, Piefke, Kartoffel oder Kraut (2). ...
Full-text available
Zusammenfassung Parallel zum Anstieg von Hate Speech in sozialen Netzwerken und anderen Online-Medien ist auch das Interesse an diesem Thema in verschiedenen Fachdisziplinen gewachsen, so dass sich mittlerweile ein eigenes Forschungsfeld etabliert hat, zu dem die unterschiedlichen Disziplinen ihren Beitrag leisten. In diesem Artikel soll der Beitrag der Linguistik zur Hate-Speech-Forschung erläutert werden, indem ein Überblick über die in diesem Feld existierende Forschung gegeben wird. Der Fokus liegt hierbei auf Untersuchungen, die dazu dienen, die sprachlichen Charakteristika von Hate Speech zu beschreiben, und weniger, diese automatisiert zu erkennen. Neben einer Zusammenfassung des Forschungsstandes ist es diesem Beitrag jedoch auch ein Anliegen, die in den verschiedenen Untersuchungen erarbeiteten Charakteristika von Hate Speech exemplarisch vorzustellen. Hierfür werden illustrierend Twitter-Daten aus drei Sprachen (Englisch, Deutsch, Niederländisch) verwendet. Der Beitrag bemüht sich jedoch auch zu zeigen, dass eine rein sprachbasierte Analyse von Hate Speech um eine multimodale Komponente erweitert werden muss, um der kommunikativen Realität der sozialen Medien ausreichend Rechnung zu tragen.
... There is no doubt that very little is known about how hate speech is perceived and evaluated by ordinary language users (Brown & Sinclair 2020); one of the rare exceptions being studies like that of Leets (2002). Therefore, as part of the XPEROHS project (Baumgarten et al. 2019), we identify typical morphosyntactic and stylistic features of hate speech in everyday written and spoken language and, moreover, analyze both qualitatively and quantitatively how these feature differences affect the perceived severity of hate-speech expressions. The perception is measured in terms of the raters' personal opinion that is transferred to two scales; one expressing the level of an utterance being personally unacceptable (henceforth called "unacceptability"), and one expressing the necessity of juridical or societal penalties for the author or speaker. ...
... 3, followed by their discussion and reflection in Sect. 4. Finally, conclusions and perspectives for future research are summarized in Sect. 5. , is grammatically and semantically annotated and accessible via a password-protected platform (Bick & Didriksen 2015). Detailed overviews of the XPEROHS corpus are provided by Baumgarten et al. (2019) and by Geyer, Bick and Kleene (in this volume). ...
This article contrastively analyzes the perception of written and spoken hate speech. The basic question is: How do linguistic features of hate-speech expressions affect the assessment of respondents with regards to people’s personal unacceptability of these expressions and their possible consequences for the author or speaker? Three specific aspects are addressed by means of an online survey. Firstly, we investigate the effect of the communication medium with written and spoken hate speech stimuli. Secondly, we compare different types of hate speech, defined by specific linguistic feature conditions as to their effects on reader/listener assessments. Thirdly, we examine how the two minority groups foreigners and Muslims that are addressed in the comments determine the perception of hate-speech expressions. The results suggest that less the medium in terms of written and spoken stimuli makes a difference in the assessment of hate-speech expressions than much more the addressed minority group as well as the linguistic features that are used by the author or speaker. We discuss our results with respect to their implications for identifying, classifying, and evaluating hate-speech expressions on social media platforms.
... While expressing their anger through intense violent remarks, some delegitimators attempted to self-portray a fair-minded identity. Likewise, legitimators also exploited the same strategy as a shield against criticism from the delegitimators (Baumgarten et al. 2019). They first expressed an understanding of the opposing argument and then presented their opinions via analogy and the discursive construction of a hypothetical future. ...
Following the first coronavirus case reported to the World Health Organization in Wuhan in 2019 and the ensuing city-wide lockdown that was imposed, many people attempted to leave the city, culminating in a vigorous discourse on the dominant Chinese microblogging site, Weibo. This study seeks to examine how online participants discursively delegitimated and legitimated people who left Wuhan before the lockdown. Weibo posts with the hashtag #逃离武汉 (‘Fleeing Wuhan’) were collected, and delegitimation and legitimation strategies deployed by users were identified. My findings reveal that the delegitimators exploited moral evaluation and impersonal authority to highlight the construed unethicality and shamelessness of people who left Wuhan, whereas the legitimators used an array of strategies, including explanation and definition, to normalize their intentions and counter linguistic hostility. These findings also provide implications vis-à-vis the clustering of delegitimation strategies as well as their linkages with emotional appeals in online discourse.
... XPEROHS is an international, German-Danish research project that examines hate speech from a cross-linguistic point of view and, most importantly, represents perhaps the first project in which the links between linguistic features and their impact on recipients are systematically investigated in perception experiments. The aim of XPEROHS is to provide social and political decision-makers with more concrete guidelines on what hate speech really is (Baumgarten et al., 2019). ...
Full-text available
The recipient is a stimulus-external factor that has so far hardly been investigated in hate-speech research. However, addressing this factor is essential to understand how and why hate speech unfolds its negative effects and which characteristics of the recipient influence these effects. The present study focuses on the recipient. Building on previous findings from explicit ratings and initial successful replications of such ratings through biosignals, we are conducting the first large-scale, systematic, and cross-linguistic biosignal study on hate speech based on two EEG measures: the beta-alpha ratio associated with arousal and the frontal alpha asymmetry associated with valence. A total of 50 Danish and German participants took part and were presented with spoken and written hate-speech stimuli, derived from authentic hate-speech posts on Twitter. Results show that Danes reacted more sensitively than Germans to hate speech containing figurative language (swear words), while Germans reacted more sensitively to hate speech with Holocaust references than Danes. In addition, teachers and lawyers showed less negative reactions to hate speech than church employees, students, and pensioners. The effect of the presentation medium depended on the respective hate speech type. In particular, speaking out hate speech based on irony and indirectness attenuated its effects on recipients to such an extent that it is questionable whether the stimuli were still perceived as instances of hate speech at all. We discuss the results in terms of key tasks of future studies and practical implication for the punishment and management of hate speech on social media.
... It must be noted that, while uniform data distribution among classes in this paper is presented as highly beneficial (from a classification point of view), it also introduces biases [13][14][15][16]. It should also be noted that, in many real-life use cases (e.g., the notion of balance of linguistic corpora [17][18][19]) a dataset that represents both classes with equal number of instances would not be representative (one class (word/phrase) can be represented with significantly less instances than other one). ...
Full-text available
This paper proposes a concept of Digital Stereotypes, observed during research on quantitative overrepresentation of one class over others, and its impact on the results of the training of Deep Learning models. The real-life observed data classes are rarely of the same size, and the intuition of presenting multiple examples of one class and then showing a few counterexamples may be very misleading in multimodal classification. Deep Learning models, when taught with overrepresentation, may produce incorrect inferring results, similar to stereotypes. The generic idea of stereotypes seems to be helpful for categorisation from the training point of view, but it has a negative influence on the inferring result. Authors evaluate a large dataset in various scenarios: overrepresentation of one or two classes, underrepresentation of some classes, and same-size (trimmed) classes. The presented research can be applied to any multiclassification applications, but it may be especially important in AI, where the classification, uncertainty and building new knowledge overlap. This paper presents specific ’decreases in accuracy’ observed within multiclassification of unleveled datasets. The ’decreases in accuracy’, named by the authors ’stereotypes’, can also bring an inspiring insight into other fields and applications, not only multimodal sentiment analysis.
... This result is in line with research by Jeung and Nham (2020) who examined the experiences of East Asian-Americans during the COVID-19 pandemic. It is possible that acts of hate speech online are less likely to be considered a criminal offence than verbal or physical harassment 'offline' , and therefore remain unreported (Baumgarten et al. 2019;Costello et al. 2019). Future research should therefore explore perceptions of the nature and severity of victimization across different settings (i.e. ...
Full-text available
We assessed whether the COVID-19 outbreak in the United Kingdom was associated with a rise in sinophobic hate crimes as well as the temporal distribution of victimization rates. A victimization survey (N = 393) showed that following the first known case of COVID-19 in the United Kingdom, Chinese/East Asian persons had a higher likelihood of being hate crime or incident victims than members of other ethnic minority groups. Specifically, victimization reported by Chinese/East Asian participants reached its highest level in March 2020 (before lockdown); it then dropped significantly after an initial relaxation of restrictions in May 2020. Overall, we documented a temporary, potentially slightly delayed hate crime trigger effect of the COVID-19 outbreak.
Hate speech continues to be an issue of key social significance, yet while its lexical and discursive aspects have been widely studied, its grammatical traits have been hitherto overlooked. This book seeks to address this gap by bringing together a global team of scholars to explore the morphosyntactic features of hateful and aggressive discourse. Drawing on thirteen diverse cross-linguistic case studies, it reveals how hate is expressed in political discourse, slang, and social media, and towards a range of target groups relating to gender, sexual orientation, and ethnic identity. Based on ideas from functional and cognitive linguistics, each thematic part demonstrates how features such as morphology, word formation, pronoun use, and syntactic structures are manipulated for the purpose of expressing hostility and hate. An innovative approach to an age-old problem, this book is essential reading for researchers and students of hate speech and verbal aggression.
Full-text available
Resumo: este artigo objetiva analisar estratégias argumentativas empregadas na veiculação de discursos de ódio contra minorias étnicas, de gênero/orientação sexual e religiosas no Facebook. Foi feita a coleta manual de comentários potencialmente ofensivos, sendo o corpus composto por 225 comentários de ódio com caráter lgbtfóbico, 194 comentários com teor racista e 181 comentários que apresentam intolerância religiosa, somando 600 comentários. Os dados foram analisados, levando-se em conta alguns trabalhos que tratam da argumentação como prática discursiva (AMOSSY, 2018; CHARAUDEAU, 2008, 2010), bem como trabalhos sobre o discurso de ódio (FORTUNA; NUNES, 2018). Ao final, apresentamos algumas categorias argumentativas para o discurso de ódio, visando à aplicação a análises periciais no âmbito da Linguística Forense e a tarefas de anotação de corpus no âmbito da Linguística Computacional. Palavras-chave: argumentação; discurso de ódio; violência verbal. Abstract: this article aims to analyze argumentative strategies used in the dissemination of hate speech against ethnic, gender/sexual orientation and religious minorities on Facebook. Manual collection of potentially offensives comments was carried out in a way that the corpus is composed by 225 comments with lgbtphobic aspect, 194 comments with racista content and 181 comments which presents possible religious intolerance, adding up 600 comments. The data was analyzed throuth works that considers argumentation as a discoursive practice (AMOSSY, 2018; CHARAUDEAU, 2008, 2010), as well as works about hate speech (FORTUNA; NUNES, 2018). In the end, we present some argumentative categories of hate speech, which can be applied to judicial expertise, in the escope of Forensic Linguistics, and to corpus annotation tasks in the ambit of Computational Linguistics. Keywords: argumentation; hate speech; verbal violence.
Full-text available
Autori u svom radu razmatraju aktualni problem zakonski reguliranog neprihvatljivog ponašanja u Hrvatskom medijskom prostoru, uključujući i društvene mreže, koncentrirajući se na problem govora mržnje. Temeljno je pitanje je li prisutnost govora mržnje u porastu. Primjenom metode analize sadržaja, a na obuhvatu kompletnog medijskog prostora Hrvatske, autori zaključuju da je porast prisutnosti neprihvatljivog ponašanja i govora mržnje hrvatska svakodnevica. Osnovom dobivenih rezultata istraživački napori autora usmjeravaju se prema detekciji vrsta govora mržnje i na njihove primjere pojavnosti pri čemu detekcija pojavnosti neprihvatljivog govora a i govora mržnje pripada vrsti usmjerenoj prema etnicitetu, religiji i izbjeglicama. U primjeni detekcije, obrade i analize su računalne tehnologije na čijem razvoju i testiranju rezultata autori nastavljaju istraživanje. Treći istraživački napor autora usmjeren je prema događajima koji su prisutni u medijskom prostoru, a odnose se na neprihvatljivi govor, te kvalitativnom i kvantitativnom analizom sadržaja daju kontekstualni okvir najvažnijih događaja.
Full-text available
The paper presents a survey study that investigates the self-conscious emotion of feeling offended and provides an account of it in terms of a socio-cognitive model of emotions. Based on the qualitative and quantitative analysis of the participants’ answers, the study provides a definition of offense and of the feeling of offense in terms of its “mental ingredients,” the beliefs and goals represented in a person who feels this emotion, and finds out what are its necessary and aggravating conditions, what are the explicit and implicit causes of offense (the other’s actions, omissions, inferred mental states), what negative evaluations are offensive and why. It also shows that the feeling of offense is not only triggered about honor or public image, but it is mainly felt in personal affective relationships. The paper finally highlights that high self-esteem may protect a person against the feeling of offense and the constellation of negative emotions triggered by it.
Full-text available
The issue of hate speech has received significant attention from legal scholars and philosophers alike. But the vast majority of this attention has been focused on presenting and critically evaluating arguments for and against hate speech bans as opposed to the prior task of conceptually analysing the term ‘hate speech’ itself. This two-part article aims to put right that imbalance. It goes beyond legal texts and judgements and beyond the legal concept hate speech in an attempt to understand the general concept hate speech. And it does so using a range of well-known methods of conceptual analysis that are distinctive of analytic philosophy. One of its main aims is to explode the myth that emotions, feelings, or attitudes of hate or hatred are part of the essential nature of hate speech. It also argues that hate speech is best conceived as a family resemblances concept. One important implication is that when looking at the full range of ways of combating hate speech, including but not limited to the use of criminal law, there is every reason to embrace an understanding of hate speech as a heterogeneous collection of expressive phenomena. Another is that it would be unsound to reject hate speech laws on the premise that they are effectively in the business of criminalising emotions, feelings, or attitudes of hate or hatred.
Conference Paper
Full-text available
Some users of social media are spreading racist, sexist, and otherwise hateful content. For the purpose of training a hate speech detection system, the reliability of the annotations is crucial, but there is no universally agreed-upon definition. We collected potentially hateful messages and asked two groups of internet users to determine whether they were hate speech or not, whether they should be banned or not and to rate their degree of offensiveness. One of the groups was shown a definition prior to completing the survey. We aimed to assess whether hate speech can be annotated reliably, and the extent to which existing definitions are in accordance with subjective ratings. Our results indicate that showing users a definition caused them to partially align their own opinion with the definition but did not improve reliability, which was very low overall. We conclude that the presence of hate speech should perhaps not be considered a binary yes-or-no decision, and raters need more detailed instructions for the annotation.
[Der findes ikke resumé til denne artikel]
This is the first comprehensive account of the Appraisal Framework. The underlying linguistic theory is explained and justified, and the application of this flexible tool, which has been applied to a wide variety of text and discourse analysis issues, is demonstrated throughout by sample text analyses from a range of registers, genres and fields.