The importance of the detection of aggressiveness in social media is due to real effects of violence provoked by negative behavior online. For this reason, hate speech online is a real problem in modern society and the necessity of control of usergenerated contents has become one of the priorities for governments, social media platforms and Internet companies. Current methodologies are far from solving this problem. Indeed, several aggressive comments are also disguised as sarcastic. In this perspective, this research proposal wants to investigate the role played by creative linguistic devices, especially sarcasm, in hate speech in multilingual context.
The web facilitates the large resonance of
hate speech, inciting racism, misogyny or
xenophobia also in the real world. Actually,
it is common that misbehaviours online are
traduced in physical attacks, such as rapes
or bulling. For instance,
In addition, the persistence and diffusion of misogynistic or offensive content can hurt and distress psychologically the victims, causing sometime their suicide, such as the case of the teenager Amanda Todd in 20121. In order to contrast the origin of these hate 1https://www.theguardian. com/commentisfree/2012/oct/26/ events and to monitor the uncontrolled flow of users texts, several initiatives have been taken in the last years. An example is the campaign No Hate Speech Movement 2 of the Council of Europe for human rights online.
The growing interest of NLP (Natural
Language Processing) research community is
demonstrated by the proposal of national
and intern
The fact that the majority of data are collected from Twitter or Facebook supports the analysis of the computer-mediated communication. As well as, the context of short text incites the creativity of authors who use figurative devices to express their opinion. One of the most used figures of speech to manifest negative opinions is the sarcasm. In fact, it is used to disguise and, at the same time, to reinforce the negative thinking, such as: i) Un pensiero di ringraziamento ogni mattina va sempre ai comunisti che ce li hanno portati fino a casa musulmani rom e delinquenti grazie7.
The ironic sharpness of the sarcasm seems
to be appropriated to express contempt and
to offend individuals subtly. In order to
study this correlation between sarcasm and
hate speech, we proposed the shared task
IronITA8
Moreover, we participated in two t
7Each morning, I would like to thank communists who bring home musulmans, roms and delinquents thanks. Tweet from IronITA corpus.
8http://di.unito.it/ironita18
9https://mexa3t.wixsite.com/home/
aggressive-detection-track
10https://
The rest of the paper is structured as follows. Section 2 introduces the literature that inspired our investigation. Section 3 describes our participation in IberEval tasks with the used approach and obtained results. In Section 4 we analyze the presence of sarcasm in analyzed aggressive and offensive texts. Finally, in Section 5 and 6 we draw our research proposal and the future work. 2
The literature about hate speech detection
includes different issues, such as:
cyberbullying, misogyny, nastiness and
aggressiveness. The most commercial methods,
currently, rely on the use of blacklists. However,
filtering the messages in this way does not
provide a sufficient remedy because it falls
short when the meaning is more subtle or
altered by sarcasm. Actually, some authors,
such as
In this context, the research is
oriented at investigating deeply the language
using classical
For the first time,
In the following section we report how we addressed the identification of aggressiveness and misogyny in Twitter, the experiments carried out and the results obtained.
Considering our motivations, our early
experiments focus mainly on hate speech
detection. For this purpose, we particip
Aggressiveness detection The first task aims to classify aggressive and non-aggressive tweets in Mexican Spanish. We applied a deep learning approach incorporating into CNN architecture a set of linguistic features (DL+FE) concerning: proper characteristics of a tweet, such as emoticons, abbreviations and slang words; stylistic information, such as the length of tweets, the use of the punctuation and the uppercase characters; bags of words weighted with tf-idf; emotive traits of the aggressiveness; and derogatory adjectives and vulgar expressions typical of Mexican culture.
By means of Information Gain, we noticed that anger and disgust are the principal emotions that incite the aggressive behaviour. We compared this system with a simple CNN architecture (DL) in order to understand the contribution of features to deep learning approach. The measure used for the competition is F-score for positive class (i.e. aggressive class). Despite the novel approach, the results obtained are low and the features seem not to help deep learning, as showed in Table 1.
DL DL+FE
0.34 0.27 Rec. 0.34 0.38 F-pos 0.34 0.31
9 10
Therefore, in order to understand what are the difficulties of DL+FE, we carried out the error analysis. We mainly noticed that there are several humorous cases, especially sarcastic (see Section 4), which are misclassified. 3.2
Automatic misogyny identification ish tweets. In the case a tweet is classified as misogynistic (Task A), we need to distinguish (Task B) if the target is an individual or not (Tar.) and identify the type of misogyny, according to the following classes (Cat.): stereotype and objectification, dominance, derailing, sexual harassment and threats of violence, and discredit. This subdivision of misogyny allows us to explore the different aspects of misogyny and compare them in two different languages. Moreover, the data are not geolocalized. Therefore, in order to gather the linguistic variations and consider the various traits of misogyny, we proposed an approach based on stylistic features captured by means of the character n-grams, sentiment and affective information, and on a set of lexicons concerning: sexuality, profanity, femininity, human body and stereotypes. In addition, we considered slangs, abbreviations and hashtags.
By means of Information Gain, we discovered some differences between the two languages: sexual language is more used in English misogynistic tweets, whereas profanities or vulgarities are more used in Spanish ones. For this task, we applied Support Vector Machine (SVM) and majority voting technique. To evaluate the Task A the organizers used Accuracy measure and for Task B the average Macro-F1 measure. In Table 2 and Table 3 we report the promising results obtained with better runs for both languages.
En Sp
Ensemble Ensemble
Acc 0.87 0.81 greur et emportement 11, that is a kind of aggressive and sharp irony addressed a target to hurt or criticize him without to exclude the possibility to amuse. This statement is corroborate by our analyses on English, Spanish, Mexican and Italian hate speech corpora. As said above, we carried out the error analysis for both tasks.
In the first competition we noticed that our approach fails in the classification of sarcastic aggressive utterances, such as: ii) @USUARIO #LOS40MeetAndGreet 9 .
su bendicio`n12.
Actually, the sarcasm is a type of figurative devices that modifies the perception of message, hindering the correct detection of hate speech by automatic systems. We found, in fact, the same difficulty for the recognition of misogynistic tweets in both languages, such as: iii) ¿Cu´al es la peor desgracia para una mujer? Parir un var´on, porque despu´es de tener un cerebro dentro durante 9 meses, van y se lo sacan13; iv) What’s the difference between a blonde and a washing machine? A washing machine won’t follow you around all day after you drop a load in it.
In virtual as in real life, sexist jokes are
very common. In general, they are considered
innocent by the majority of people.
However,
These early observations suggest the necessity to address the use of figures of speech such as sarcasm, in order to accurate, in 11“type of irony done with sharpness and a fit of anger”
12@User #LOS40MeetAndGreet 9 . Because she is a fighter mother who takes care of her kid.
13What’s the worst disgrace for a woman? Giving birth to boy, because after she has got a brain into her for 9 months, it is taken out a multilingual perspective, the automated methods to flag abusive language.
For this purpose, we propose an accurate analysis of different kinds of hate speech online especially in Italian, English and Spanish, taking into account also the geographical linguistic variations. We focus in particular on short texts such as tweets, posts or comments, exploring the informal language.
Considering the previous observations, we propose approaching the hate speech detection issues taking into account the figurative dimension of language and especially of abusive language. Moreover, it is necessary to examine the appropriateness of various computational techniques to solve this problem. In this line, we want to examine the contribution of the linguistic features to deep learning approaches by comparison with the performances of classical techniques. Finally, the multilingual context allows to discover the typical aspects of hate speech in order to recognize it independently from the languages.
Indeed, the scope of this investigation is to propose a methodology for monitoring correctly the user-generated contents allowing the system to work as sensor of the violence, also in real world. 6
Our research aims to explore the several
dimensions of hate speech considering, above
all, the use of figurative devices that hinder
the automatic processes of recognition. In
order to investigate the remarks observed in
these first experiments, as future work, we
would like to participate in HaSpeeDe14 and
AMI15
In addition, similar tasks are proposed at
SemEval 2019 concerning: multilingual hate
speech against immigrants and women
(HatEval)16, and the identification and
categorization of offensive language in social
media (OffensEval)17. Analyzing different kinds
of abusive language allows to understand the
boundaries between them and their singular
aspects. Finally, multilingual context gives
us the opportunity to delineate the
differ14http://www.di.unito.it/~tutreeb/
haspeede-evalita18/index.html\#
15https://
17https://competitions.codalab.org/ competitions/20011 ences and analogies between the various languages, inferring general characteristics of hate speech online.