The paper describes the Web platform built within the project “Contro l'odio”, for monitoring and contrasting discrimination and hate speech against immigrants in Italy. It applies a combination of computational linguistics techniques for hate speech detection and data visualization tools on data drawn from Twitter. It allows users to access a huge amount of information through interactive maps, also tuning their view, e.g., visualizing the most viral tweets and interactively reducing the inherent complexity of data. Educational courses for high school students and citizenship has been developed which are centered on the platform and focused on the deconstruction of negative stereotypes against immigrants, Roma, and religious minorities, and on the creation of positive narratives.
Hate Speech (HS) is a multi-faceted phenomenon with countless nuances, a high degree of individual and cultural variation, and intersections with related concepts such as offensive language, threats, bullying and so on.
The detection of HS is a recent yet popular task that is gaining the attention of the NLP community but also that of public institutions and private companies. There are several problems connected
Copyright c 2019 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0). with this delicate task: a cultural-dependent definition, a highly subjective perception, the need to remove potentially illegal contents quickly from the Web and the connected risk to unjustly remove legal content, the partly overlapping linguistic phenomena that make it hard to identify HS. English social media texts are the most studied, but other languages, sources and textual genres are investigated as well.
“Contro l’odio”1 is a project for countering and
preventing racist discrimination and HS in Italy,
in particular focused against immigrants. On the
one hand, the project follows and extends the
research outcomes emerged from the ‘Italian Hate
Map project’
This paper focuses on the technological core of the project, a Web platform that combines computational linguistics analysis with visualization techniques, in order to provide users with an inter1https://controlodio.it/ 2http://acmos.net/ 3http://www.voxdiritti.it/ active interface for exploring the dynamics of the discourse of hate against immigrants in Italian social media. Three typical targets of discrimination related to this topical focus are taken into account, namely Migrants, Muslims and Roma, since they exemplify discrimination based on nationality, religious beliefs and ethnicity, respectively. Since November 2018 the platform analyses daily Twitter posts and exploits temporal and geo-spatial information related to messages in order to ease the summarization of the hate detection outcome. 2
In the last few years several works contributed to
the development of HS detection automatic
methods, both releasing novel annotated resources,
lexicons of hate words or presenting automated
classifiers. Two surveys were recently published on
this topic
Since 2016, shared tasks on the detection of
HS or related phenomena (such as abusive
language or misogyny) in various languages have
been organized, benefiting from the developed
datasets and effectively enhancing advancements
in resource building and system development both.
These include in particular HatEval at SemEval
2019
The project “Contro l’odio” follows and extends
the research outcome emerged from the ‘Italian
Hate Map project’
3.1
The architecture consists of four main modules. The data collection module gathers the tweets by using the Stream Twitter API and filters them by keywords. The automatic classifier module automatically annotates the presence of HS in the filtered tweets, relying on a supervised approach. The next module stores the annotated tweet aggregating them by time and place in a database. The last module, implemented by relying on a node.js server, exposes the API that are requested by the front end (Figure 1). 3.2
We started collecting tweets from October 1st 2018 by using the Twitter’s Stream API. The
6http://hatemeter.eu/ 7http://mandola-project.eu/ 8http://www.antiatlas.net/geographyof-hate-geotagged-hateful-tweets-inthe-united-states-en/ streaming is filtered using the vowels as keywords and the alpha-2 code it as language filter. About 700; 000 Italian statuses are daily gathered, but only about 17; 000 are relevant for monitoring discrimination and HS against immigrants in Italy. We filtered relevant tweets by using the keywords proposed in Poletto et al. (2017), considering three typical targets of discrimination — namely migrants, Roma and religious minorities. 3.3
In order to automatically label the tweets, we
developed a supervised classifier to predict the
presence of HS in text. The classification is binary
(i.e., presence of HS vs. absence of HS). We
employ a Support Vector Machine (SVM)
classifier with one-hot unigram representation as
feature vector. We train the classifier on the
Italian Hate Speech Corpus
The following tweets are two examples of annotated tweets: 1. #dallavostraparte non ci sono moderati, sono tutti terroristi pronti a tagliarci la testa e per questo io li odio a morte! #onyourside there are no moderates, they all are terrorists ready to cut our head off and for this I hate them to death! 2. Tanto con il sole i nomadi non vengono piu` a scuola. Per qualcuno questa e` la soluzione...
Nomads no longer come to school when it’s sunny. For some this is the solution...
In example 1, the target is “religious minorities”
and the author spreads and incites violence against
Islamic people (the tweet contains hate speech). In
example 2, the target is “Roma”, and the
previous conditions are not detected, there’s not hate
speech. By performing cross-validation
experiments on such corpus, we estimate the best
hyperparameters for the model: 27,642 features,
learning rate optimal, linear kernel. With this
settings, we record a prediction performance in
crossvalidation of 0.81 (0.70 for the class hate speech)
precision and 0.81 (0.67 for the class hate speech)
recall (Favg = 0:80 0:01). Recently, new
classification strategies base on language
understanding models have been demonstrated to be suitable
for the task of hate speech detection, obtaining
encouraging results. As a consequence, we are
considering the possibility to compare our model with
a classifier base on AlBERTo
It is important to note that the Italian Hate
Speech Corpus has been collected in a specific
time frame, from October 1st, 2016 to April 25th,
2017
The main view of the dashboard is a choropleth map and allows the user to explore the spatial
dimension (regional and provincial level) of the dataset. The temporal dimension can be explored thanks to a time slider that also shows the trend of the total number of tweets and the percentage of HS. In figure 5 there’s an example of how the choropleth map and the Dorling map appear on June 29, 2019 when the migrant and NGO themes, in a single day, become viral in the public debate10. In Figure 4 we see another example: the volume of tweets about the Roma topic in the days from 3 to 5 June 2019 has increased considerably due to some clashes in the outskirts of Rome11.
10http://www.ansa.it/sito/notizie/ politica/2019/06/28/sea-watch-indagatala-capitana.-nuovo-affondo-di-salvinicontro-lolanda-comportamento-disgustoso_ 991189d6-7818-48d9-b4d8-a2a7d10d31bc. html
11http://www.ansa.it/sito/notizie/ cronaca/2019/04/04/simone-ilquindicenne-di-torre-maura-controcasapound-state-a-fa-leva-sulla-rabbiaThe liquid gauge allows the user to quickly detect the tweet volume increase, from 1; 619 to 14; 778, and the increase in HS rates, from 13% to 23%. 4.2
Figure 2 shows another visualization: a bar chart containing the 25 words more frequently occurring in the tweets collected in the selected time period. For each word, the user can also see the average percentage of HS in tweets containing that word. As before, the example in figure 2 refers to June 29, 2019. By clicking on a word, the user can visualize additional information about it, such as the exact number of occurrences in the tweets or its co-occurrence network (figure 3).
The interactive hate maps and the ‘Words of Hate’ visualization settings described here have been also used within educational paths developed for citizenship and mostly targeting high school students. Such paths were focused on the dismantling of negative stereotypes against immigrants, Roma, and religious minorities, and on the creation of positive narratives to actively counteract hatred online. Since today, a team of twenty educators carried out 90 laboratories in seven different Italian regions (Piedmont, Tuscany, Liguria, Emilia della-gente.-plauso-raggivideo_7a4bc495bb4d-4c21-a1f7-2ecbc8422ea5.html
Romagna, Lazio, Friuli-Venezia Giulia, and Sardinia). At the end of the project 150 classes will be reached, and the resulting positive narratives will be published on the project website. 6
In this paper we described an online platform for monitoring HS against immigrants in Italy at different levels of granularity, which uses Twitter as data source and combines HS detection and advanced visualization techniques in order to provide users with an interactive interface for the exploration of the resulted data. Another important research outcome of the project is HATECHECKER, a tool that automatically detects hater users in online social networks, which will be accessible from the platform soon. Given a target user, the workflow that is going to be implemented in our system uses sentiment analysis techniques to identify hate speech posted by the user, and exploits a lexicon-based approach to assign to the person one or more labels that describe the nature of the hate speech she posted (e.g., racism, homophobia, sexism, etc.). A map of Italian projects and associations that spread a culture of tolerance is also under development, to allow ‘Contro l’Odio’ users to get a better understanding of the HS phenomenon and of the active forces fighting it on the Italian territory.
The work of all the authors was partially
funded by Italian Ministry of Labor
Anna Schmidt and Michael Wiegand. 2017. A Survey on Hate Speech Detection using Natural Language Processing. In Proceedings of the Fifth International Workshop on Natural Language Processing for Social Media. Association for Computational Linguistics.