=Paper=
{{Paper
|id=Vol-2006/paper024
|storemode=property
|title=Hate Speech Annotation: Analysis of an Italian Twitter Corpus
|pdfUrl=https://ceur-ws.org/Vol-2006/paper024.pdf
|volume=Vol-2006
|authors=Fabio Poletto,Marco Stranisci,Manuela Sanguinetti,Viviana Patti,Cristina Bosco
|dblpUrl=https://dblp.org/rec/conf/clic-it/PolettoSSPB17
}}
==Hate Speech Annotation: Analysis of an Italian Twitter Corpus==
https://ceur-ws.org/Vol-2006/paper024.pdf
Hate Speech Annotation:
Analysis of an Italian Twitter Corpus
Fabio Poletto Marco Stranisci Manuela Sanguinetti,
Dipartimento di StudiUm Acmos Viviana Patti,
University of Turin marco.stranisci@acmos.net Cristina Bosco
f.poletto91@gmail.com Dipartimento di Informatica
University of Turin
{msanguin,patti,bosco}@di.unito.it
Abstract distinguish it from offline communication and
make it potentially more dangerous and hurtful
English. The paper describes the develop- (Ziccardi, 2016). What is more, HS is featured
ment of a corpus from social media built as a complex and multifaceted phenomenon, also
with the aim of representing and analysing because of the variety of approaches employed
hate speech against some minority groups in attempting to draw the line between HS and
in Italy. The issues related to data collec- free speech (Yong, 2011). Therefore, despite
tion and annotation are introduced, focus- the multiple efforts, there is yet no universally
ing on the challenges we addressed in de- accepted definition of HS.
signing a multifaceted set of labels where From a juridical perspective, two contrasting
the main features of verbal hate expres- approaches can be recognised: while US law is
sions may be modelled. Moreover, an oriented, quite uniquely, towards granting free-
analysis of the disagreement among the dom of speech above all, even when potentially
annotators is presented in order to carry hurtful or threatening, legislation in Europe and
out a preliminary evaluation of the data set the rest of the world tends to protect the dignity
and the scheme. and rights of minority groups against any form of
Italiano. L’articolo descrive un corpus di expression that might violate or endanger them.
testi estratti da social media costruito con Several European treaties and conventions ban
il principale obiettivo di rappresentare ed HS: to mention but one, the Council of European
analizzare il fenomeno dell’hate speech ri- Union condemns publicly inciting violence or
volto contro i migranti in Italia. Vengono hatred towards persons or groups defined by
introdotti gli aspetti significativi della rac- reference to race, colour, religion, descent or
colta ed annotazione dei dati, richia- national or ethnic origin. The No Hate Speech
mando l’attenzione sulle sfide affrontate Movement1 , promoted by the Council of Europe,
per progettare un insieme di etichette che is also worth-mentioning for its efforts in endors-
rifletta le molte sfaccettature necessarie ing responsible behaviours and preventing HS
a cogliere e modellare le caratteristiche among European citizens.
delle espressioni di odio. Inoltre viene
presentata un’analisi del disagreement tra The main aim of this paper is at introducing
gli annotatori allo scopo di tentare una a novel resource which can be useful for the in-
preliminare valutazione del corpus e dello vestigation of HS in a sentiment analysis perspec-
schema di annotazione stesso. tive (Schmidt and Wiegand, 2017). Providing that
among the minority groups targeted by HS, the
present socio-political context shows that some of
1 Introduction them are especially vulnerable and garner constant
attention - often negative - from the public opin-
Hate is all but a new phenomenon, yet the global
ion, i.e. immigrants (Bosco et al., 2017), Roma
spread of Internet and social network services
and Muslims, we decided to focus our work on HS
has provided it with new means and forms of
against such groups. Furthermore, providing the
dissemination. Online hateful content, or Hate
spread of HS in social media together with their
Speech (HS), is characterised by some key aspects
1
(such as virality, or presumed anonymity) which https://www.nohatespeechmovement.org
current relevance in communication, we focused 3 Dataset Collection
on texts from Twitter, whose peculiar structure and
conventions make it particularly suitable for data The dataset creation phase was divided into three
gathering and analysis. main stages.
We first collected all the tweets written in Italian
and posted from 1st October 2016 to 25th April
2 Related Work 2017.
Then we discussed in order to establish a) which
One of the earlier attempts to develop a corpus- minority groups should be identified as possible
based model for automated detection of HS on the HS targets, and b) the set of keywords associated
Web is found in Warner and Hirschberg (2012): with each target, in order to filter the data col-
the authors collect and label a set of sentences lected in the previous step. As for the first as-
from various websites, and test a classifier for pect, we identified three targets that we deemed
detecting anti-Semitic hatred. They observe that particularly relevant in the present Italian scenario;
HS against different groups is characterised by a based also on the terminology used in European
small set of high frequency stereotypical words, Union reports2 , the targets selected for our corpus
also stressing the importance of distinguishing HS were immigrants (class: ethnic origin), Muslims
from simply offensive content. (class: religion), and Roma. As regards the sec-
The same distinction is at the core of Davidson et ond aspect mentioned above, we are aware of the
al. (2017), where a classifier is trained to recog- limits of a keyword-based method in HS identifi-
nise whether a tweet is hateful or just offensive, cation (Saleem et al., 2016), especially regarding
observing that for some categories this difference the amount of noisy data (e.g. off-topic tweets)
is less clear than for others. that may result from such method; on the other
An exhaustive list of the targets of online hate is hand, the choice to adopt a list of explicitly hateful
found in Silva et al. (2016), where HS on two words3 may prevent us from finding subtler forms
social networks (Twitter and Whisper) is detected of HS, or even just tweets where a hateful message
through a sentence structure-based model. is expressed without using a hate-related lexicon.
One of the core issues of manually labelling HS With this in mind, we then filtered the data by re-
is the reliability of annotations and the inter- taining a small set of neutral keywords associated
annotator agreement. The issue is confronted by with each target. The keywords selected are sum-
Waseem (2016) and Ross et al. (2017), who find marised below:
that more precise results are obtained by relying ethnic group religion Roma
on expert rather than amateur annotations, and that immigrat* terrorismo rom
the overall reliability remains low. The authors (immigrant*) (terrorism) (roma)
suggest that HS should not be considered as a bi- immigrazione terrorist* nomad*
nary ”yes/no” value and that finer-grained labels (immigration) (terrorist*) (nomad*)
may help increase the agreement rate. migrant* islam
An alternative to lexicon-based approaches is sug- stranier* mussulman*
gested in Saleem (2016), where limits and biases (foreign) (muslim*)
of manual annotation and keyword-based tech- profug* corano
niques are pointed out, and a method based on (refugee*) (koran)
the language used within self-defined hateful web
communities is presented. The method, suitable The dataset thus retrieved consisted of 370,252
for various platforms, bypasses the need to define tweets about ethnic origins, 176,290 about religion
HS and the inevitable poor reliability of manual
annotation. 2
See the 2015 Eurobarometer Survey on discrimination
While most of the available works are based on in the EU: http://ec.europa.eu/justice/
fundamental-rights/files/factsheet_
English language, Del Vigna et al. (2017) is the eurobarometer_fundamental_rights_2015.
first work on a manually annotated Italian HS cor- pdf
3
pus: here the authors apply a traditional procedure Such as the ones extracted for the Italian HS map (Musto
et al., 2016):
on a corpus crawled from Facebook, developing http://www.voxdiritti.it/ecco-la-nuova-
two classifiers for automated detection of HS. edizione-della-mappa-dellintolleranza/
and 31,990 about Roma. target tweet
The last stage consisted in the creation of the religion Ci vuole la guerra per salvare l’Italia
corpus to be annotated. In order to obtain a bal- dai criminali filo islamici
anced resource, we randomly selected from the (”We need a war to save Italy from
previous dataset 700 tweets for each target, with pro-Islamic criminals”)
a total amount of 2,100 tweets. In case even just one of these conditions is not
However, a large number of tweets were further detected, HS is assumed not to occur.
removed from the corpus, during the annotation In line with this definition, we also attempted
stage (because of duplicates and off-topic con- to extend the scheme to other annotation cate-
tent). Despite the size reduction, though, the dis- gories that seemed to significantly co-occur with
tribution of the targets in the corpus remained HS; this in order to better represent the (perceived)
quite unchanged, resulting in a balanced resource meaning of the tweet, and to help the annotator in
in this respect. the task, by providing a richer and finer-grained
At present, the amount of annotated data con- tagset4 . The newly-introduced categories are de-
sists of 1,828 tweets. In the next section, we scribed below.
describe the whole annotation process and the
scheme adopted for this purpose. Aggressiveness (labels no - weak - strong): it fo-
cuses on the user intention to be aggressive, harm-
4 Data Annotation: Designing and ful, or even to incite, in various forms, to violent
Applying the Schema acts against a given target; if the message reflects
an overtly hostile attitude, or whenever the target
Being HS a complex and multi-layered concept, group is portrayed as a threat to social stability,
and being the task of its annotation quite difficult the tweet is considered weakly aggressive, while
and prone to subjectivity, we undertook some pre- if there is the reference – whether explicit or just
liminary steps in order to make sure that all anno- implied – to violent actions of any kind, the tweet
tators share a common ground of basic concepts, is strongly aggressive.
starting from the very definition of HS. tweet aggressiveness
When determining what can, or cannot, be consid- nuova invasione di migranti weak
ered HS (thus in a yes-no fashion), and based on in Europa
the juridical literature and observations reported (A new migrant invasion in Europe)
above in Section 1, we considered two different
factors: Cacciamo i rom dall’Italia strong
(Let’s kick Roma out of Italy)
• the target involved, i.e. the tweet should be
Offensiveness (labels no - weak - strong): con-
addressed, or just refer to, one of the minority
versely to aggressiveness, it rather focuses on the
groups identified as HS targets in the previ-
potentially hurtful effect of the tweet content on
ous stage (see Section 3), or even to an indi-
a given target. A tweet is considered weakly of-
vidual considered for its membership in that
fensive in a large number of cases, among these:
category (and not for its individual character-
the given target is associated with typical human
istics);
flaws (laziness in particular), the status of disad-
vantaged or discriminated minority is questioned,
• the action, or more precisely the illocution- or when the members of the target group are de-
ary force of the utterance, in that it is capable scribed as unpleasant people; on the other hand, if
of spreading, inciting, promoting or justify- an overtly insulting language is used, or the target
ing violence against a target. is addressed to by means of outrageous or degrad-
ing expressions, the tweet is expected to be con-
Whenever both factors happen to co-occur in sidered as strongly offensive.
the same tweet, we consider it as a HS case, as in 4
The whole scheme description along with the de-
the example below: tailed guidelines are available at https://github.com/
msang/hate-speech-corpus
tweet offensiveness Annotation process The annotation task con-
I migranti sanno solo weak sisted in a multiple-step process, and it was carried
ostentare l’ozio out by four independent annotators after a prelimi-
(Migrants can only show off nary step where the guidelines were discussed and
their laziness) partially revised.
The corpus was split in two, and each part was
Zingari di merda strong annotated by two annotators. The annotator pairs
(You fucking Roma) then switched to the other part, in order to provide
a third (possibly solving) annotation to all those
Irony (labels no - yes): it determines whether
tweets where at least one category was labelled
the tweet is ironic or sarcastic rather than based on
differently by the previous two annotators. A fur-
the literal meaning of words. The introduction of
ther subset of around 130 tweets still received dif-
this category in the scheme was led by preliminary
ferent labels by the different annotators (namely
observations of the data, which highlighted how it
for aggressiveness and offensiveness). In order to
was a fairly common linguistic expedient used to
solve these remaining cases, a fifth independent
mitigate or indirectly convey a hateful content.
annotator was finally involved. As a result, the
tweet irony final corpus only contains tweets that were fully
ora tutti questi falsi profughi yes revised.
li mandiamo a casa di Renzi ??! Regarding the results of the annotation in terms
(shall we send all these of label distribution, we found that 16% of all
fake refugees to Renzi’s house??!) tweets have been considered containing HS, of
which 23% against immigrants, 38% against Mus-
Stereotype (labels no - yes): it determines lims and 39% against Roma. When considered
whether the tweet contains any implicit or ex- alone, aggressiveness occurs in 14% , offensive-
plicit reference to (mostly untrue) beliefs about a ness in 10%, irony in 11% and stereotype in 29%
given target. There is a whole host of stereotypes of tweets. However, the labels that co-occur more
and prejudices associated with the target groups frequently with hate speech are those indicating
selected for our research; from an exploratory the presence of aggressiveness (81%), stereotypes
observation of the data in the corpus, the fol- (81%), and offensiveness (56%), and, overall, they
lowing cases were identified: the members of a co-occur altogether 52% of the times; irony is la-
given target are referred to as invaders, freeload- belled in 11% of HS tweets. While, within the
ers, criminals, filthy (or having filthy habits), sex- whole corpus, 57% of cases are just tweets with a
ist/mysoginist, undemocratic, violent people. “neutral” content, which means that no one of the
Furthermore, we also take into account the role categories were annotated as such.
that conventional media may have in spreading
stereotypes and prejudices while reporting news 4.1 Agreement Analysis
on refugees, migrants, and minorities in general. The development phase related to the inter-
Based on what suggested in the Italian journalists’ annotator agreement (IAA) is not only a necessary
Code of Conduct, known as ”Carta di Roma”5 , in step for validating the corpus and evaluating the
order to ensure a correct and responsible reporting schema adopted, but also a tool that provides more
about these topics, we also applied this criterion to details about the trends and biases of individual
any tweet containing a news headline that implic- annotators with respect to specific annotation cat-
itly endorses, or contributes to the spread of, such egories.
stereotypical portrayals (see the example below). In this study, we measured the IAA right after
the first annotation step was completed, i.e. the
tweet stereotype
one where just two annotators were involved (see
Roma in bancarotta ma regala yes
Section 4). In line with related cases6 , our data
12 milioni ai rom
showed a very low agreement: in 47% of cases,
(Rome is bankrupt but still gives
the annotator pair annotated at least one of the five
12 millions to Roma)
6
See (Del Vigna et al., 2017), (Gitari et al., 2015), (Kwok
and Wang, 2013), (Ross et al., 2017), (Waseem, 2016), to
5
See https://www.cartadiroma.org/ mention a few.
categories using different labels. In fact, the dis- 5 Conclusion and Future Work
agreement took place mostly in one (40%) or two
(16%) categories, while just 4 tweets received a We introduced in this paper the collection and an-
completely different annotation by the annotator notation of an Italian Twitter corpus representing
pairs. More specifically, we measured the agree- HS towards some selected target. Our main aim is
ment coefficient, using Cohen’s kappa (Carletta, at producing a corpus to be used for training and
1996), for each individual category. Results – also testing sentiment analysis systems, but some effort
reported in Table 1 – show that the category with must still be applied to achieve this goal. The cur-
the highest agreement is namely the one related to rent contribute is mainly in designing and trying a
the presence of hate speech (abbreviated to ‘hs’ in novel schema for HS, but the relatively low agree-
the table), followed by irony (‘iro.’) and stereotype ment shows that modelling this phenomenon is a
(‘ster.’). very challenging task and a further refinement of
the guidelines and of the scheme must be applied,
together with the application to larger data sets.
hs aggr. off. iro. ster. Acknowledgments
before merge 0,54 0,18 0,32 0,44 0,43
after merge 0,54 0,43 0,37 0,44 0,43 The work of Cristina Bosco, Viviana Patti
and Manuela Sanguinetti was partially funded
Table 1: Agreement (Cohen’s k) for each annota- by Progetto di Ateneo/CSP 2016 (Immigrants,
tion category before and after merging labels for Hate and Prejudice in Social Media, project
aggressiveness and offensiveness. S1618 L2 BOSC 01) and partially funded by
Fondazione CRT (Hate Speech and Social Media,
project n. 2016.0688).
Considering that the lowest agreement was
found in aggressiveness (‘aggr.’) and offensive- References
ness (‘off.’) – the only categories where three la-
Cristina Bosco, Patti Viviana, Marcello Bogetti,
bels were used, instead of two – the agreement was
Michelangelo Conoscenti, Giancarlo Ruffo,
recalculated by merging the weak-strong labels; it Rossano Schifanella, and Marco Stranisci. 2017.
thus increased considerably (especially in aggres- Tools and resources for detecting hate and prejudice
siveness), though still remaining far below an ac- against immigrants in social media. In Proceedings
ceptable threshold. of First Symposium on Social Interactions in Com-
plex Intelligent Systems (SICIS), AISB Convention
The low agreement with regard to the degree of 2017, AI and Society, Bath, UK.
offensiveness can be attributed to the absence of
clear indications within the annotation guidelines Jean Carletta. 1996. Assessing agreement on classi-
in this respect. fication tasks: The kappa statistic. Computational
Linguistics, 22(2):249–254, June.
Finally, among the annotation criteria established
in the preliminary stage, one in particular proved Thomas Davidson, Dana Warmsley, Michael Macy,
to be quite misleading, i.e. whenever a clearly and Ingmar Weber. 2017. Automated hate speech
hateful tweet did not actually refer to the target detection and the problem of offensive language. In
International AAAI Conference on Web and Social
identified by one of the selected keywords, HS
Media.
and stereotype were assumed not to occur. On the
other hand, the remaining categories should be an- Fabio Del Vigna, Andrea Cimino, Felice Dell’Orletta,
notated accordingly. This principle was conceived Marinella Petrocchi, and Maurizio Tesconi. 2017.
in order to provide annotated data that could be Hate me, hate me not: Hate speech detection on
Facebook. In Proceedings of the First Italian Con-
considered a true reflection of HS towards the tar- ference on Cybersecurity (ITASEC17), Venice, Italy,
gets we identified in our study, though still ”pre- January 17-20, 2017., pages 86–95.
serving” the meaning and the intent of the tweet in
itself, regardless of the target involved. This, to- Njagi Dennis Gitari, Zhang Zuping, Hanyurwimfura
Damien, and Jun Long. 2015. A lexicon-based
gether with other points of the guidelines, will be approach for hate speech detection. International
further discussed and clarified in the next project Journal of Multimedia and Ubiquitous Engineering,
phase. 10(4):215–230.
Irene Kwok and Yuzhou Wang. 2013. Locate the hate:
Detecting tweets against blacks. In Marie desJardins
and Michael L. Littman, editors, AAAI. AAAI Press.
Cataldo Musto, Giovanni Semeraro, Marco de Gem-
mis, and Pasquale Lops. 2016. Modeling com-
munity behavior through semantic analysis of so-
cial data: The italian hate map experience. In Pro-
ceedings of the 2016 Conference on User Modeling
Adaptation and Personalization, UMAP 2016, Hali-
fax, NS, Canada, July 13 - 17, 2016, pages 307–308.
Björn Ross, Michael Rist, Guillermo Carbonell, Ben-
jamin Cabrera, Nils Kurowsky, and Michael Wo-
jatzki. 2017. Measuring the reliability of hate
speech annotations: The case of the European
refugee crisis. CoRR, abs/1701.08118.
Haji Mohammad Saleem, Kelly P. Dillon, Susan Be-
nesch, and Derek Ruths. 2016. A web of hate:
Tackling hateful speech in online social spaces. In
Proceedings of the First Workshop on Text Analyt-
ics for Cybersecurity and Online Safety (TA-COS
2016)s, Portoro, Slovenia.
Anna Schmidt and Michael Wiegand. 2017. A survey
on hate speech detection using natural language pro-
cessing. In Proceedings of the Fifth International
Workshop on Natural Language Processing for So-
cial Media, pages 1–10, Valencia, Spain, April. As-
sociation for Computational Linguistics.
Leandro Silva, Mainack Mondal, Denzil Correa, Fab-
rcio Benevenuto, and Ingmar Weber. 2016. Ana-
lyzing the targets of hate in online social media. In
Proceedings of the 10th International Conference on
Web and Social Media, ICWSM 2016, pages 687–
690. AAAI Press.
William Warner and Julia Hirschberg. 2012. Detecting
hate speech on the world wide web. In Proceedings
of the Second Workshop on Language in Social Me-
dia, LSM ’12, pages 19–26, Stroudsburg, PA, USA.
Association for Computational Linguistics.
Zeerak Waseem. 2016. Are you a racist or am i seeing
things? annotator influence on hate speech detection
on Twitter. In Proceedings of the First Workshop
on NLP and Computational Social Science, pages
138–142, Austin, Texas, November. Association for
Computational Linguistics.
Caleb Yong. 2011. Does freedom of speech include
hate speech? Res Publica, 17(4):385, Jul.
Giovanni Ziccardi. 2016. L’odio online. Violenza ver-
bale e ossessioni in rete. Saggi / [Cortina]. Cortina
Raffaello.