=Paper=
{{Paper
|id=Vol-3033/paper35
|storemode=property
|title=Recognizing Hate with NLP: The Teaching Experience of the #DeactivHate Lab in Italian High Schools
|pdfUrl=https://ceur-ws.org/Vol-3033/paper35.pdf
|volume=Vol-3033
|authors=Simona Frenda,Alessandra Teresa Cignarella,Marco Antonio Stranisci,Mirko Lai,Cristina Bosco,Viviana Patti
|dblpUrl=https://dblp.org/rec/conf/clic-it/FrendaCSLBP21
}}
==Recognizing Hate with NLP: The Teaching Experience of the #DeactivHate Lab in Italian High Schools==
https://ceur-ws.org/Vol-3033/paper35.pdf
Recognizing Hate with NLP: The Teaching Experience of the
#D EACTIV H ATE Lab in Italian High Schools
Simona Frenda1,2 , Alessandra Teresa Cignarella1,2 , Marco Antonio Stranisci1 , Mirko Lai1 ,
Cristina Bosco1 and Viviana Patti1
1. Università degli Studi di Torino, Italy
2. Universitat Politècnica de València, Spain
{simona.frenda|alessandrateresa.cignarella|marcoantonio.stranisci|mirko.lai}@unito.it
{cristina.bosco|viviana.patti}@unito.it
Abstract these effects might be suicide, especially consider-
ing the adolescents, as suggested by recent studies
The possibility of raising awareness about investigating the link between cyberbullying and
misbehaviour online, such as hate speech, suicidal behaviors of U.S. youth (Nikolaou, 2017).
especially in young generations could To prevent such scenarios, few awareness-raising
help society to reduce their impact, and projects in schools are activated by NGOs in Italy,
thus, their consequences. The Com- such as Amnesty International1 or Cifa ONLUS2 .
puter Science Department of the Univer- The Commissione Orientamento e Informat-
sity of Turin has designed various tech- ica nelle scuole3 supports a manifold of activ-
nologies that support educational projects ities with the main goal of creating a link be-
and activities in this perspective. We tween schools and academia, also in the con-
implemented an annotation platform for text of the national project Piano Lauree Scien-
Italian tweets employed in a laboratory tifiche (PLS). The members of the CCC (Content-
called #D EACTIV H ATE, specifically de- Centered Computing) group of the Computer Sci-
signed for secondary school students. The ence Department of the University of Turin, active
laboratory aims at countering hateful phe- in the investigation of hate speech online4 , have
nomena online and making students aware led and participated in several hate-speech-related
of technologies that they use on a daily ba- projects, including “Contro l’odio”5 (Capozzi et
sis. We describe our teaching experience al., 2020) a joint work with non-profit entities and
in high schools and the usefulness of the University of Bari that aims at monitoring hate
technologies and activities tested. speech against minorities in Italy. Within the cur-
rent experience, we created a data annotation plat-
1 Introduction form specifically dedicated to support educational
activities and aimed at reflecting on the impor-
Recently, the presence of digital technologies in tance of a conscientious communication. In this
our lives has grown enormously, with a strong perspective, the idea of #D EACTIV H ATE takes
impact on our daily lives. Digital spaces and hold. This laboratory, addressed at students of sec-
social media have become a privileged channel ondary schools, is articulated in three main mod-
for communication, information and socialization, ules with the purpose of:
frequented by millions of people at the same time.
Along with the new relational opportunities and 1) raising awareness about this social problem,
access to knowledge, even misbehaviour have ac- encouraging the reflection on microaggres-
quired new visibility and virality, such as hate sions, hate speech, stereotypes, prejudices;
speech. In spite of a causal link between hate
speech and crime is hard to demonstrate, the risk 2) stimulating the so-called computational
of offences and effects on psychological and phys- thinking and the study of linguistic elements
ical well-being of the victims are clear in psy- that are exploited by users to offend or to ex-
chological and social studies (Nadal et al., 2014; 1
http://di.unito.it/silencehateitaly.
Fulper et al., 2014). The extreme consequence of 2
http://di.unito.it/iorispetto.
3
Copyright © 2021 for this paper by its authors. Use per- http://di.unito.it/orientamentoscuole.
4
mitted under Creative Commons License Attribution 4.0 In- http://hatespeech.di.unito.it/.
5
ternational (CC BY 4.0). https://controlodio.it/.
press hate against a victim online (hashtags, students’ opinions, is somehow in continuity with
emoticons, or figures of speech); that experience.
A second work of great relevance for the cre-
3) introducing high schoolers to how tools based ation of our experience, has been the reading of
on NLP (Natural Language Processing) work Pannitto et al. (2021), in which the authors point
to incentivize a more conscious use of tech- out, for the first time, the fact that no high school
nology. curricula in Italy includes any (computational) lin-
guistics education and that the lack of this kind of
To reach these purposes, We designed a series of
exposure makes choosing computational linguis-
educational activities that include: analysis of the
tics as a university degree unlikely. Furthermore,
online problem by means of an investigation on
the authors highlight that NLP is, indeed, at the
own social networks personal profiles; linguistic
core of many tools young people use in their ev-
analysis of the hateful messages during the anno-
eryday life, and having almost zero knowledge of
tation of tweets on the “Contro l’odio” annotation
this field makes the use of such tools less responsi-
platform; manual identification of hate speech in
ble than it could be. The authors have been the first
Italian texts playing the role of ‘being an auto-
to create a dedicated workshop for Italian, aimed
matic classifier’; translation of this task in a real
at raising awareness of Italian students aged be-
automatic task, coding two types of classifiers in
tween 13 and 18 years regarding the subject of
Python. These activities, delivered online due to
NLP (Messina et al., 2021).
the pandemic restrictions, have been distributed in
Additionally, the idea of creating some play-
5 meetings (lasting 2 hours each) for each class,
ful and meaningful activities regarding NLP and
between April and June 2021, for a total of 10
the themes of hate speech for high schoolers, are
hours per class.
in line with the concept of ‘gamification’, which
2 Related Work lately has been applied to many linguistic annota-
tion tasks, as an alternative to crowdsourcing plat-
A popular workshop series on the topic of “Teach- forms to collect annotated data in an inexpensive
ing NLP” has been recently held on its fifth edition way (Bonetti and Tonelli, 2020), such as our “Con-
at NAACL-HLT 2021 (Jurgens et al., 2021), where tro l’odio” annotation platform.
the participants discussed and shared experiences
on a variety of important issues such as: teaching 3 #D EACTIV H ATE
guidelines, teaching strategies, adapting to differ-
ent student audiences, resources for assignments, The goals of #D EACTIV H ATE are: 1) raising
and course or program design. The main lesson awareness about misbehaviour online, such as
learned has been that of highlighting the impor- hate speech, eliciting also personal experiences, 2)
tance of creating materials describing NLP, not stimulating computational thinking and linguistic
only for learners at a university/college level, but observation of hateful messages, and 3) encour-
also for those learners who are younger and have aging a conscious use of technologies discovering
diverse educational backgrounds. In this regard, a how they work. To reach these objectives we ar-
great inspiration for starting to work with schools ticulated three modules as described below.
in Italy derives from the experience of Sprugnoli et
3.1 Hate Speech: Introduction
al. (2018), where the authors – although with dif-
ferent goal in mind than ours – started a project in- The first module aims at introducing a definition
volving NLP and pupils from Italian schools, aged of hate speech to students. Hate speech is often
12-13. That experience was chiefly dedicated to mistaken for a generic insult rather than a specific
the study of cyberbullying among pre-teens and phenomenon “connected with hatred of members
the creation of a corpus of WhatsApp threads in of groups or classes of persons identified by cer-
the context of the CybeRbullying EffEcts Preven- tain ascriptive characteristics (e.g., race, ethnicity,
tion activities (CREEP) project. Our idea of start- nationality)” (Brown, 2015).
ing a project that could bring NLP to high school- The session started with an ice-breaking activity
ers and that, at the same time, could introduce the in which students presented themselves through an
themes of hate speech, microaggressions, and dis- image found online, depicting an aspect of their
crimination by eliciting personal experiences and identity (see Figure 1). We then asked them to tell
whether they were ever attacked or stigmatized for After a brief introduction on what corpora are
this characteristic. and how they are used in new technologies, stu-
dents have been involved in an annotation task of
hate speech, asking them to evaluate at least 30
tweets.
For this purpose, we created the data annota-
tion platform8 within the “Contro l’odio” project.
This web application, built using PHP, MySQL,
and JavaScript, 9 , preserves the student’s annota-
tion history by using a passwordless authentica-
tion link sent to the email chosen during the login.
Figure 1: Example of Jamboard of Google This method has the twofold advantage of not re-
quiring the student to register to the platform and
of preventing ourselves to save the student’s email
In this way, we guided the class in drawing a
or other personal data. It then ensures the anno-
distinction between non-ascriptive identity traits
tation anonymity and satisfies the requirements of
(e.g., political belief, style of dressing) and ascrip-
General Data Protection Regulation (GDPR), as a
tive6 ones (e.g., ethnicity, sexual orientation, skin
desired consequence.
colour) (Reskin, 2005). The idea behind this activ-
The home page of the web application consists
ity is twofold: i) it links issues such as hate speech
of a dashboard that provides the annotation guide-
and racial microaggression (Sue, 2010) to stu-
line and shows basic information about the stu-
dents’ lives; ii) it helps distinguishing the spread-
dent’s activity. Indeed, the student could know the
ing of discriminatory contents7 from generic in-
number of sessions they completed (each session
sults. The module ended with an assignment: stu-
consists of annotating 15 tweets) and the level of
dents had to find at least one public figure who had
agreement (expressed in percentage) between their
been a victim of online discrimination, providing
annotation and the annotation performed by the
one or more hateful messages as an example, and
automatic model realized in the “Contro l’odio”
a counter-narrative response.
project. Gamifying the task through this compar-
3.2 “If I Were a Classifier...” ison, we provide the basis for a discussion about
the fallibility of automatic systems. Furthermore,
The second module is organized in two meetings we also allow the student to compare their annota-
and focuses on the importance of manually anno- tion with the annotation of their classmates in or-
tated corpora for online hate speech detection and der to introduce the measures of annotator agree-
what are the peculiarities of hateful messages. ment. When a session starts, the student could an-
Within the first meeting, each student presented notate the level of hatefulness of a tweet through a
the found messages and try to define the type of 7 square scale filled with a color scale from Watusi
attack and the linguistic characteristics of the text to Sangria as shown in Figure 2. Two additional
that make it hateful or a counter-narrative. The squares, respectively filled with White and Mid-
variety of examples led to the introduction of a Gray, allow stating the absence of hate or to con-
deeper taxonomy of discrimination (e.g., misog- sider off-topic the content of the tweet. Finally,
yny, homophobia, sexism, etc...). As expected, the three toggle switches (on/off button) were added
following group discussion brought out a consid- to check the presence of ‘irony/sarcasm/humor’,
erable subjectivity in perceiving these phenomena, ‘offensiveness’, and ‘stereotype’, giving them the
thus highlighting the need of adopting a shared an- possibility to reflect about the ways in which users
notation schema to identify hate speech in mes- spread hate online.
sages. During the annotation task, students were asked
6
Qualities beyond the control of an individual.
to fill a shared spreadsheet with the tweets that im-
7
The definition of hate speech we referred to is the one pressed them the most for its offensiveness, for its
codified by The Council of Europe: “the term ‘hate speech’ humorous intention, or the most difficult to anno-
shall be understood as covering all forms of expression which
8
spread, incite, promote or justify racial hatred, xenophobia, https://didattica.controlodio.it/.
9
anti-Semitism or other forms of hatred based on intolerance” https://github.com/mirkolai/DEACTIVH
(Recommendation No. R (97) 20). ATELab.
most controversial tweets report aggressive events
or racial propositions; and, for this reason, they
were perceived as hurtful by the majority of the
students:
(i) Autobus per i bianchi e altri per i migranti. Non si
parla dell’apertheid del Sudafrica né del periodo di
segregazione negli Stati Uniti, ma di una proposta della
Lega per la provincia di Bergamo. L’Italia non è un
paese razzista ma nel 2020 questo è ciò di cui si dis-
cute. URL10
Others triggered interesting linguistic reflections,
such as:
(ii) Peccato che non sbarcano povere famiglie africane, ma
solo mafia nigeriana, ex galeotti tunisini, stupratori
Figure 2: Data Annotation Platform senegalesi, terroristi dell’Isis dalla libia, tutti crimi-
nali robusti 1.80 di altezza, pronti a spacciare droga,
violentare le nostre donne, cannibali e assassini.11
tate. By discussing with them annotation results,
In these, the students retrieved specific figures of
we introduced the latest core concept of the mod-
speech such as sarcasm, rhetorical questions and
ule: the agreement. We presented some metrics
analogies, and also strong words that reflect the
that are typically adopted to calculate it among an-
social biases towards the minorities. In activity A,
notators and outlined some good practice recently
all the words and expressions that could make the
emerged in Corpus Linguistics, such as ensuring
message hurtful have been collected in a list of n-
the involvement of minorities in corpora develop-
grams of words called our lexicon (Table 1).
ment in order to avoid biases (Basile, 2020).
Following, the items of such list have been ex-
3.3 My First Classifier ploited by the classifiers to predict if a tweet con-
tains hate speech or not.
In this module the main idea is to stimulate com-
putational thinking by translating linguistic ob- unigrams risorse, sporchi, pacchia, schifo, inva-
servations coming from the annotation procedure sione, spacciare
n-grams porti chiusi, cacciarli via, difesa della
in a proper computational task. The activity of patria12
annotation has, indeed, given the opportunity of
reflecting on how users tend to verbally express Table 1: Examples from our lexicon
hate online, and on how minorities are represented
through stereotypes. To incentivize this transition, For activity B, we created an interactive Python
we proposed two activities: notebook using the Colaboratory platform pro-
vided by Google, as a similar initiative had suc-
A. to mark in each tweet the textual span that cessfully been carried out by Hiippala (2021) with
could make a classifier aware of the pres- a similar educational tool. To allow the students
ence of hate speech creating a list of word to use the notebook in spite of their computer
n-grams; skills, we elaborated some guidelines explaining
even how to create a folder in Google Drive and
B. to develop two automatic classifiers (super-
10
vised and unsupervised) exploiting the list of Translation: Buses for whites and others for migrants.
There is no mention of South Africa’s apartheid or the period
word n-grams. of segregation in the United States, but of proposal by Lega
for the province of Bergamo. Italy is not a racist country but
Before starting with the first activity, we asked in 2020 this is what we are discussing. URL.
11
students to motivate their choice of the tweets se- Translation: Too bad that poor African families do not
land, but only the Nigerian mafia, former Tunisian convicts,
lected during the previous exercise. Some tweets Senegalese rapists, ISIS terrorists from Libya, all heavy-
triggered a discussions on what should be consid- weight criminals 1.80 tall, ready to sell drugs, rape our
ered hate speech or not, and the doubts were later women, cannibals and murderers.
12
Translation: Unigrams: resources, dirty, godsend, dis-
solved by looking at the provided definitions of gust, invasion, peddle. N-grams: closed harbours, send
hate speech and at the annotation guidelines. The [them] away, defence of the fatherland.
how to import all the necessary materials inside of making available the necessary materials to stu-
it. Among the required materials, we prepared the dents. Moreover, each meeting was supported by
dataset using the tweets previously annotated by the use of slides for having visual and descriptive
the students. support. The classes assisted in this short period
We proposed two types of classifiers: were composed of a total of 35 adolescents, com-
ing from different countries. From the first meet-
1) unsupervised classifier based on the list ing they showed a general interest in the treated
our lexicon for which if one of the se- subject, and we were surprised especially by the
lected grams are inside the text, the text is profoundness of some observations raised during
predicted as hateful; the discussions. The students, indeed, were en-
2) supervised classifier based on Support couraged to share their opinions, doubts, and per-
Vector Machine algorithm using the list spectives. These discussions made clear that the
our lexicon as main feature of the students face these problems related to technology
classification task. and communication every day, sometimes suffer-
ing even the consequences. Hate speech is, indeed,
The coding of the first classifier allowed students a very sensitive issue and the perception of what
to gain confidence with some basics of Python; is abusive or not, depends on the cultural back-
whereas the second one introduced them to core ground of each student. This fact, on the one side
of new technologies based on machine learning stimulated the debates, however, on the other side,
(see Figure 3). At the end of the activity, we it made it difficult for us to find the ideal way to
observed together the performances of automatic share complex concepts and manage specific situ-
systems and analyzed some of the tweets that were ations.
wrongly classified. This final step helped stu- At the end of the laboratory, we provided a sur-
dents to reflect on the limitations of machines and vey in order to collect the impressions and the
the important role of the linguistics in language- opinions of students. Analyzing these surveys, we
related technologies. noticed that the majority of students considered
interesting the content of #D EACTIV H ATE, but it
4 What We Learnt appears clear that the format online of the labora-
Due to pandemic restrictions, we taught the entire tory was perceived from students less interactive
laboratory through remote modality (DAD)13 be- and fluent, due especially to technical problems
tween April and June 2021 to 2 classes of one sec- when a part of students were in class and other part
ondary school of Turin, with students aged 16-20. at home14 . From our perspective, we noticed an
As described above, various resources and tools interesting difference between younger and older
have been used (and created ex novo) to bring for- students. The older were more active during the
ward the educational activities in distant teaching activities and discussions than the younger. More-
mode. However, we plan to propose the same ac- over, we thought that the number of students af-
tivities/materials even for lessons in presentia ex- fected the flow of the debates, especially in the
ploiting the computer rooms of the schools. DAD context. We expect that in presentia the pro-
For each class, we organized the activities of posed activities could have a better impact facili-
the three modules in 5 meetings of about 2 hours. tating the interaction.
Despite the shortness of the laboratory, we found
5 Conclusion
that realizing specific activities for each session
helped us manage efficiently the available time. #D EACTIV H ATE represents for Italian high
We resorted to web applications to make up for schoolers a first step towards the introduction to
the different devices and operating systems used subjects such as Linguistics and NLP, that are, for
by the students at their homes. And, in particular, the most part, unknown in Italian high schools, in
we used Google Meet, as it offers interactive tools spite of their relevance in everyday technology. In-
such as virtual blackboard, and Moodle, a learning deed, this kind of laboratory reveals what are the
platform provided by the University of Turin that possible hybrid and multidisciplinary applications
gave us the possibility to organize our activities
14
For the most part of the school year 2020-2021, Italian
13
Didattica A Distanza. schools allowed a capacity of 50% inside classrooms.
Figure 3: Supervised Classifier Section on Python Notebook
of Computer Science and Linguistics related de- tomatic text processing, as well as a final evalua-
grees, far from the conventional employment op- tion of the proposed teaching activities collecting
portunities. Looking at the future, we would like the personal impressions of the students.
to enhance the proposed activities in order to make In addition, to validate also the impact of #D E -
them more interactive even in an online context ACTIV H ATE in the society and, in particular, in the
(such as the DAD) following the example of Hiip- city context we think to measure the detection of
pala (2021). the amount of hateful message online by means of
A final remark needs to be made regarding the monitoring platforms, such as the “Contro l’odio”
lack of evaluative strategies that could allow us map.15
to understand the impact of #D EACTIV H ATE in
students’ online behaviors or their knowledge of Acknowledgements
technologies. Therefore, following the example
The work of S. Frenda, A. T. Cignarella and M.
of Bioglio et al. (2018) and Athanasiades et al.
Lai has been funded under the national project
(2015), in the next editions we have planned to
Piano Lauree Scientifiche (PLS) 2019/20 as part
employ: surveys before and after the interven-
of the activities of Computer Science Department,
tion to evaluate the online activity of the students
School of Science of Nature, University of Turin.
and their experiences about misbehavior (caused
The authors would like to extend a special thanks
or suffered); and interviews to teachers after the
to the school ‘Convitto Nazionale Umberto I’, and
conclusion of the laboratory to understand if some
in particular, to Professor Simona Ventura for her
changes were perceived with respect to the class
availability and her collaboration in this adventure
group. Future activities will integrate also basic
with #D EACTIV H ATE.
evaluations to assess the degree of learning with
respect to the contents of the course, such as com-
15
putational thinking, annotation methodologies, au- https://mappa.controlodio.it/.
References of racial microaggressions on mental health: Coun-
seling implications for clients of color. Journal of
Christina Athanasiades, Harris Kamariotis, Anastasia Counseling & Development, 92(1):57–66.
Psalti, Anna C Baldry, and Anna Sorrentino. 2015.
Internet use and cyberbullying among adolescent Dimitrios Nikolaou. 2017. Does cyberbullying impact
students in Greece: the “Tabby” project. Hellenic youth suicidal behaviors? Journal of health eco-
Journal of Psychology, 12(1):14–39. nomics, 56:30–46.
Valerio Basile. 2020. It’s the End of the Gold Standard Ludovica Pannitto, Lucia Busso, Claudia Roberta
as We Know It. In International Conference of the Combei, Lucio Messina, Alessio Miaschi, Gabriele
Italian Association for Artificial Intelligence, pages Sarti, and Malvina Nissim. 2021. Teaching NLP
441–453. Springer. with Bracelets and Restaurant Menus: An Interac-
tive Workshop for Italian Students. In Proceedings
Livio Bioglio, Sara Capecchi, Federico Peiretti, Den- of the Fifth Workshop on Teaching NLP, Online. As-
nis Sayed, Antonella Torasso, and Ruggero G Pensa. sociation for Computational Linguistics.
2018. A social network simulation game to raise
awareness of privacy among school children. IEEE Barbara F. Reskin. 2005. Including mechanisms in
Transactions on Learning Technologies, 12(4):456– our models of ascriptive inequality. Handbook of
469. employment discrimination research, pages 75–97.
Federico Bonetti and Sara Tonelli. 2020. A 3D Rachele Sprugnoli, Stefano Menini, Sara Tonelli, Fil-
Role-Playing Game for Abusive Language Annota- ippo Oncini, and Enrico Piras. 2018. Creating a
tion. In Workshop on Games and Natural Language WhatsApp Dataset to Study Pre-teen Cyberbullying.
Processing, pages 39–43. European Language Re- In Proceedings of the 2nd Workshop on Abusive Lan-
sources Association. guage Online (ALW2), pages 51–59. Association for
Computational Linguistics.
Alexander Brown. 2015. Hate speech law: a philo-
sophical examination. Routledge. Derald Wing Sue. 2010. Microaggressions in every-
day life: Race, gender, and sexual orientation. John
Arthur TE Capozzi, Mirko Lai, Valerio Basile, Fabio Wiley & Sons.
Poletto, Manuela Sanguinetti, Cristina Bosco, Vi-
viana Patti, Giancarlo Ruffo, Cataldo Musto, Marco
Polignano, et al. 2020. “Contro L’Odio”: A
Platform for Detecting, Monitoring and Visualizing
Hate Speech against Immigrants in Italian Social
Media. IJCoL. Italian Journal of Computational
Linguistics, 6(6-1):77–97.
Rachael Fulper, Giovanni Luca Ciampaglia, Emilio
Ferrara, Y Ahn, Alessandro Flammini, Filippo
Menczer, Bryce Lewis, and Kehontas Rowe. 2014.
Misogynistic language on Twitter and sexual vio-
lence. In Proceedings of the ACM Web Science
Workshop on ChASM.
Tuomo Hiippala. 2021. Applied Language Technol-
ogy: NLP for the Humanities. In Proceedings of the
Fifth Workshop on Teaching NLP, pages 46–48, On-
line, June. Association for Computational Linguis-
tics.
David Jurgens, Varada Kolhatkar, Lucy Li, Margot
Mieskes, and Ted Pedersen, editors. 2021. Pro-
ceedings of the Fifth Workshop on Teaching NLP.
Association for Computational Linguistics.
Lucio Messina, Lucia Busso, Claudia Roberta Combei,
Alessio Miaschi, Ludovica Pannitto, Gabriele Sarti,
and Malvina Nissim. 2021. A Dissemination Work-
shop for Introducing Young Italian Students to NLP.
In Proceedings of the Fifth Workshop on Teaching
NLP, pages 52–54, Online, June. Association for
Computational Linguistics.
Kevin L Nadal, Katie E Griffin, Yinglee Wong, Sahran
Hamit, and Morgan Rasmus. 2014. The impact