While favouring communications and easing information sharing, Online Social Networks are increasingly used to launch harmful campaigns against speci c groups and individuals. Although providers struggle to keep pace by manually removing hate content published on their platforms, recent research e orts rely on automatic text classi cation techniques, whose performances are usually measured on annotated corpora. In this work, we propose six distinct machine learning classi cation strategies: three based on conventional machine learning approaches, three based on neural networks. The latter are able to process texts almost from scratch, avoiding the need of i) NLP tools specialised for a speci c language, ii) the phase of time-consuming feature engineering, and iii) the high computational cost usually derived from processing a huge amount of features. Thus, the main goal of the paper is to investigate whether it is possible to rely on neural networks and to achieve performance results at least comparable with those of NLP-based classi ers. The performances of the six con gurations are evaluated over an annotated dataset consisting of 4,000 Italian tweets and 4,000 Italian Facebook comments. By comparing the classi cation results, we demonstrate that relying on deep learning techniques for hate speech detection is more than encouraging. In particular, a deep learning model, based on an ensemble approach, obtains a F1 score of 0.786 on the Twitter data and 0.775 on the Facebook ones, the best results, compared to the ones obtained with the other tested con gurations.
The more and more massive usage of Online Social Networks (OSNs) leads
undeniable bene ts, among others the opportunity for users to easily interact for
a myriad of goals, which range from planning social events to engaging in
commercial transactions. Unfortunately, since years social platforms also represent a
fertile ground for ill-intentioned people, whose goals spam from maliciously
inuencing the public opinion, by di using polarized content on important societal
topics { like politics, terrorism, immigration { (see, e.g., [
Furthermore, the enormous degree of freedom for knowledge creation and
sharing on OSNs allows the publication of content promoting `violence or
hatred against individuals or groups based on certain attributes, such as: race or
ethnic origin, religion, disability, gender, age, veteran status, sexual orientation
and gender identity'2. Such violent content is usually referred as hate speech.
Although the most popular social networks managers announced in their policies
to strive to ght against hate speech [21], current solutions mainly rely on
manually checking and possibly removing the targeted content once published, upon
appropriate signaling by OSNs users. Despite the massive enrollment of content
moderators by social media providers [31], the recent past has seen the dramatic
growth of hate attacks, a ecting the mental and even the physical status of the
victims (see [
A promising stream of research to ght haters, by quickly and automatically discriminating between hate and no hate speech, is the training of automated classi ers based on manually annotated corpora. The majority of the existing methods rely on supervised document classi cation tasks [29]. In their turn, these tasks are divided into two main categories: classic and deep learning methods [35]. While the former depend on the extraction and engineering of (mainly) textual features, successively given as input by classi ers such as Support Vector Machines, Naive Bayes, and Logistic Regression, see, e.g., [7], the latter employ neural networks to learn various level of abstract features from raw text [24]. As of late, the trend in automatic detection of hate speech is relying on deep learning. This is mainly due to the fact that classic methods strongly depend on Natural Language Processing (NLP) approaches, which require a notable e ort for extracting the features in input to the classi er and, moreover, are strongly dependent from the considered language.
The research question of this work is the following: Can we avoid the adoption of NLP-based classi ers for the task of hate speech detection, moving towards deep-learning techniques and achieving at least comparable performance results?
To answer the question, we compare the performances of NLP and deep learning-based classi ers to automatically discriminate whether, and to what extent, two Italian corpora express hate. Our benchmark consists of two annotated datasets: i) a collection of Facebook posts and comments, created and rstly used in [8], and ii) a Twitter corpus recently introduced in [26, 28]. 1 https://www.mobistealth.com/blog/facebook-misuse-stats/ All URLs accessed on November 16, 2018. 2 https://support.google.com/youtube/answer/2801939?hl=en YouTube Hate Speech Policy.
Overall, we test six classi er con gurations: three based on conventional machine learning approach, three based on neural networks. The outcome of the analysis, given as usual in terms of standard classi cation metrics, testify that the best classi cation results are obtained via deep learning techniques. This means that an e ective hate speech detection system can be successfully built without the need of NLP tools. The main advantages of this achievement are: { a language-agnostic solution: there is no need of NLP tools optimised for speci c languages; { a very limited feature engineering phase: basically, we use tokenization and stop words removal only; { a sizable reduction of the time needed for both the learning and the classication phase: at the same classi cation performances, deep learning-based classi ers need a number of features which are 3 orders of magnitude less than those needed by NLP-based classi ers; { there is no need for a complete re-training of the learning model, when feeding the classi er with new labeled data: neural networks models can be updated online, and incrementally, on the arrival of new data. Remarkably, this limits the issue known as concept drift [11].
As a further achievement, although for both the Twitter and the Facebook datasets the best results are obtained with the deep learning methodologies, we discuss the di erences of such results passing from one methodology to the other. This gives us some remarkable insights on the characteristics of the corpora taken into consideration, and consequently, the capability to discern how much is the relative gain in the use of each methodology for a speci c dataset.
Remainder: Next section describes the annotated corpora and how the text is pre-processed and, next, it introduces the six classi cation con gurations adopted throughout the paper. Section 3 describes the choice of parameters of the learning methods and critically discusses the obtained performances. Finally, Section 4 illustrates related work in the area of hate speech detection, while Section 5 concludes the paper. 2
We propose a comparison between several methods of increasing complexity which can be used to build an e ective hate speech detection system. The main aim is to show that we can detect hate speech by completely avoiding the \feature engineering problem" (adopting textual representations based on language modeling techniques [16]) and using popular deep learning algorithms instead. Each of the proposed methods is a binary classi er which outputs 1 in case of an input document expressing hate and 0 otherwise. Before describing each method, we introduce the dataset used throughout this study, and we describe how we process documents before applying machine learning methods to them. 2.1
The dataset is the result of a joint e ort of two research teams on unifying the
annotation previously applied to two di erent datasets, in order to allow their
exploitation for the HaSpeeDe (Hate Speech Detection) task [
The rst sub-dataset is a collection of Facebook comments created in 2016 and rstly presented in [8]. The content of the comments to Facebook posts were retrieved through a versatile Facebook crawler, which exploited the Facebook Graph API4. The collected comments are related to posts published on a series of public Italian web pages and groups, mainly involved in politics, and possibly featuring hate content. The original set in [8] consisted of 17,567 Facebook comments from 99 posts.
The second sub-dataset is a Twitter corpus [26, 28] comprising tweets against immigrants, Muslims and Roma. To obtain such data, a phase of data ltering was previously enacted, by means of a keyword-based approach (neutral keywords frequently associated to the three targets). This led to 236,193 tweets. After further processing and restricting the dataset so obtained, the nal version of the corpus was made of 6,928 tweets.
Recently, the Facebook and Twitter corpora have been re-annotated and made publicly available for participating to the Evalita 2018 task on Hate Speech Detection. In the brand new annotated dataset, the annotation format consists of a simpli ed version of the schemes adopted in [8, 26, 28]. The resulting annotated dataset comprises the tweet or Facebook comment along with the tag resulting from the annotation, 1 and 0, expressing the presence or not of hate speech in the text. Overall, the renewed Facebook and Twitter dataset consists of 4,000 comments and 4,000 tweets. Each corpus is divided into two distinct sets (3,000 documents for training and 1,000 for test). For Facebook, as training set we have 1,618 comments tagged as 0 (No Hate) and 1,382 tagged as 1 (Hate), while as test set we have 323 comments tagged as 0 and 677 tagged as 1. For Twitter, as training set we have 2,028 tweets tagged as 0 and 972 tagged as 1, while as test set we have 654 tweets tagged as 0 and 346 tagged as 1. 2.2
We use two di erent approaches for the textual representations of documents, i.e., bag of words and word embeddings (more on these later). Before processing the data with machine learning methods, we tokenize the original text of the Twitter and Facebook input datasets, by adopting the following approach. After lowering the text and removing the stopwords, we delete all the URLs. Next, we include all tokens which belong to one of these categories: a) normal words, b) emoticons, and c) special characters (e.g., `!' and `?'). In particular, we remove 3 http://www.di.unito.it/~tutreeb/haspeede-evalita18/index.html\# 4 https://developers.facebook.com/docs/graph-api `#' from hashtags and consider the remaining characters as single words, while we include all the usernames i.e., (words beginning with `@') as single tokens.
The rst textual representation we use is the classic bag-of-words (BoW)[30], which is probably the most popular approach used in the past years. With this representation, we analyze the training data to build a dictionary of valid words (tokens) and use them to encode the documents by generating high sparse vectors for their representations. The importance of each token contained in a document is usually weighted according to a speci c policy. Here, we adopt the well known TF-IDF [30], whose main idea is to give more importance to terms that are frequent in a document included in the corpus under evaluation, but, at the same time, tend to appear in few documents.
The second textual representation is based on word embeddings, a set of neural language modeling techniques[12] based on unsupervised learning and used to build e ective vector representations of textual contents. Here, we consider the popular word2vec algorithm[23] implemented in gensim software5 as the language modeling technique to represent text. This technique extracts lowdimensional dense word vectors from analyzed text which keep track of semantic/syntactic relationships existent between words. Such vectors can thus be used in algebraic operations to point out some speci c characteristic of the data (e.g., W (\queen") = W (\king") W (\man") + W (\woman")).
We use three di erent word2vec models to encode documents. The rst one is a model built over the complete document collection of the Italian Wikipedia6, using skipgram architecture, a window equals to 10 words and embeddings size equals to 300. The main problem of this model is that the Wikipedia articles used to train it are very di erent in terms of lexical and syntactical forms from documents available on input datasets (i.e., short texts like tweets and comments on Facebook). To overcome this limitation, we also consider a second embeddings model, which is based on the rst one but it is updated considering the documents collection coming from the Twitter dataset. Thirdly, we build a third word2vec model based on the Wikipedia one but updated with data coming from the Facebook dataset.
In order to encode the documents, we compute the document vector di as lwi di = 1=lwi X we(idx(i; j))
j=1 where lwi is the number of valid words7 included in document i, idx(i; j) is the function which returns the index of embeddings vector corresponding to the word at position j in document i, and we is the function which returns the embeddings vector corresponding to the speci ed index. This document encoding is desirable while analyzing short textual documents, such as in our scenario, because the resulting vector does not su er from a di use information loss, due to the mixing and averaging of vectors of many words (typical case for long documents). 5 A Python implementation of word2vec: https://radimrehurek.com/gensim/. 6 The pre-trained embeddings model: http://hlt.isti.cnr.it/wordembeddings/. 7 A word is valid only if it is available as entry in the dictionary of embeddings model. 2.3
In the following we present the methods tested usable for building e ective hate speech detection systems. First we present 3 methods based on SVM algorithm but using di erent textual data representations. Next we describe 3 di erent deep learning algorithms, of increasing complexity, which show very competitive results and avoid almost completely the feature-engineering process, except for the pre-processing part previously described.
As rst method (called `SVM-TfIdf' in Table 2), we use a classi er based on the SVM algorithm[14]8. This algorithm has been successfully used in several works related to text classi cation[15, 32, 10], especially if combined with BoW textual representations and TF-IDF feature weighting. SVM, even when used with default parameters (e.g., with a simple linear kernel) usually performs reasonably well on various text classi cations tasks. For this reason, we argue that this algorithm, together with BoW and TF-IDF, could be a quite strong baseline to compare with, giving us evidence of how the other tested methods perform.
To improve the e ectiveness of the algorithm, we use other two variants of SVM, by combining it with a textual representation based on embeddings, obtained as described in Section 2.2. The rst variant (`SVM-GenEmb' in Table 2) encode documents using a generic embeddings model built over the Italian Wikipedia collection. A second variant (`SVM-SpeEmb' in Table 2) try to take advantages from a specialized version of the previous embeddings model, enriched with the Facebook and Twitter datasets. We therefore generate two embeddings models, one speci c for Twitter and one for the Facebook data. (a) CNN architecture (b) GRU architecture
(c) Ensemble architecture
8 SVM implementation provided by scikit-learn, http://scikit-learn.org/stable/.
We subsequently concentrate on deep learning (DL) methods, which, in the last years, have proven to perform very well on text classi cation tasks. The rst architecture (referred as `CNN' in Table 2) is shown in Figure 1a. The core is a Convolutional Neural Network (CNN) [17], a fast and popular DL method, often used for image classi cation tasks, but also proven e ective within NLP domains [34]. The proposed architecture encodes a raw text into a list of token IDs, which translates in activating the corresponding tokens vectors into an embeddings layer, pre-loaded before that the training starts, with weights learned in the embeddings model9. The encoded document is then fed to a CNN layer that tries to learn some space-invariant high level features, able to catch interesting patterns for nal hate speech detection phase.
CNN does not take into account the order and dependencies of words while
learning potential interesting features. Thus, the same identi ed pattern is
given the same weight, independently from contextual dependencies and
position within the document. To overcome this limitation, we test a second DL
architecture ( Figure 1b). It includes a Gated Recurrent Unit (GRU) layer [
The last tested method, referred as `Ensemble' in Table 2 and Figure 1c, tries to merge advantages from the previous DL architectures. Here, there is an ensemble classi er whose decisions are based on 3 di erent sub-modules. The rst two modules are composed by two separated layers, as discussed above (GRU and CNN networks). The third one is a composition of layers, rst a CNN layer and then a GRU layer. The main intuition for such a con guration is that the CNN and GRU single sub-modules could produce independent information about a text, to be classi ed according to the speci c nature of the network (spatial for CNN, temporal for GRU), while the third sub-module could produce a high level and dependent spatial-temporal representation of the text. The contribution of each sub-module is in turn sent to an ensemble layer, which can appropriately measure each feedback, before taking the nal decision. 3
This section describes the assumptions and parameters used to perform the experimentation and presents and discusses the experimental results. 3.1
For all the SVM methods, except for the `SVM-TfIdf' con guration, we use a kfold cross validation (with k = 5), in order to nd those kernels and parameters 9 We always use the specialized variant of embeddings model. that guarantee the best e ectiveness on a validation set opportunely extracted from the training set10. In the case of `SVM-T df', for which the computational cost necessary to perform the optimization is very high11, we decide to use the default SVM parameters, i.e., a linear kernel with C equals to 1. In all the deep learning architectures, we use a learning procedure including a batch size of 32 documents, a maximum input sequence length of 100 words, opportunely padded if needed, a maximum number of epochs equals to 2012, a xed dropout value of 0.2 and the binary cross-entropy loss function.
A summary of the parameters values is in Table 1. The values C and Gamma are the standard parameter names used in the SVM algorithm for tuning its behaviour during the learning phase, NumFiltersCNN is the number of convolutional units used in the corresponding layer, KernelSizeCNN is the size of the sliding window used in convolutions, and NumUnitsGRU is the number of GRU units used in the corresponding layer. 3.2
The experimental results are in Table 2. We use the Precision (Pr), Recall (Re) and F1 metrics [30] as the standard way to measure the e ectiveness of the proposed methods.
We report in bold the best F1 values obtained for the two datasets, for the single classes (No hate and Hate) and for the average F1 obtained weighting both the labels according to the distribution of the documents in the two classes. 10 The training set has been divided in two distinct sets. The rst 90% used as real training in the optimization process, the remaining 10% as validation data used to measure the e ectiveness of the built model. 11 The document vector is very sparse and long, in the order of tens of thousands distinct words. 12 As early stopping criterion, in order to prevent over tting, we stop the learning process when the loss on validation set is less than 0.02 after two consecutive epochs.
Table (2) E ectiveness of tested methods over the test sets.
No hate Hate Weighted avg results
Pr Re F1 Pr Re F1 Pr Re F1 SVM-TfIdf 0.805 0.875 0.838 0.702 0.564 0.625 0.771 0.774 rSVM-GenEmb 0.810 0.870 0.839 0.679 0.574 0.622 0.767 0.774 teSVM-SpeEmb 0.817 0.867 0.841 0.682 0.596 0.636 0.773 0.779 itw CNN 0.798 0.935 0.861 0.788 0.506 0.617 0.795 0.796 T GRU 0.830 0.851 0.840 0.671 0.636 0.653 0.778 0.781
Ensemble 0.839 0.848 0.843 0.675 0.660 0.668 0.786 0.787 SVM-TfIdf 0.520 0.782 0.624 0.861 0.663 0.749 0.750 0.701 okSVM-GenEmb 0.609 0.638 0.623 0.823 0.805 0.814 0.754 0.751 bSVM-SpeEmb 0.602 0.669 0.633 0.833 0.789 0.810 0.758 0.750 o ce CNN 0.636 0.681 0.658 0.843 0.814 0.828 0.776 0.771 aF GRU 0.658 0.619 0.638 0.823 0.846 0.835 0.770 0.773
Ensemble 0.647 0.659 0.653 0.836 0.829 0.832 0.775 0.774 Considering the weighted average results, the baseline method SVM-TfIdf works quite well, especially for Twitter, where it obtains good performances, comparable to the SVM methods using textual representations with embeddings (F1 equal to 0.769, 0.769, 0.775, resp.). Instead, on Facebook, the baseline seems to su er (F1 equal to 0.708), probably because of a higher average length of the texts which negatively in uences the discriminating power of IDF part in feature weighting: this decreases the e ectiveness of the classi er. Still on Facebook, the two SVM methods using embeddings (SVM-GenEmb and SVM-SpeEmb) provide useful additional information from the analysis of Wikipedia data and outperform the baseline SVM classi er by a notable margin (F1 = 0.752, 0.753).
The double use of embeddings is comparable in e ectiveness, with a slight advantage of SVM-SpeEmb on Twitter data (0.775 vs 0.769), where it can exploits local information like emoticons, commonly used in tweets to express emotions.
GRU and CNN show very similar weighted average performances, always better than those obtained with the SVM-based methods. However, they present some di erences between each other. In particular, on both datasets, CNN works a lot better than GRU on documents labelled as No hate, while GRU works features good performances in discriminating documents labelled as Hate.
The Ensemble architecture takes advantages from the two DL methods and obtains a compromise generally able to guarantee the best weighted average F1 (0.786 for Twitter, 0.775 for Facebook). Indeed, even if the method does not obtain the best F1 results on all classes and datasets (with the exception of Hate on Twitter, where F1 = 0.668), it is always close to the best performer (0.843 vs 0.861 for No Hate on Twitter, 0.653 vs 0.658 for No Hate on Facebook, 0.832 vs 0.835 for Hate on Facebook). Therefore, the Ensemble con guration results in a always-good solution for the application scenarios considered in this work.
Table (3)
Method
SVM-TfIdf SVM-SpeEmb
Ensemble
Learning time (s) Classi cation time (s) 422.41 32.56 18.62 1.19 138.59 3.58
Table 3 shows how the methods perform wrt learning and classi cation phases on Facebook. For the sake of simplicity, we report only three methods, since the others add no additional information to the discussion. The measures consider both the time to build a classi er with a set of already xed algorithm's parameters (`Learning time') and the time to classify all the documents in the test set (`Classi cation time'). SVM-TfIdf is the most costly method because, with a BoW sparse representation, it must process tens of thousands of di erent features. With embeddings (SVM-SpeEmb), the computational cost is drastically decreased. The most complex tested method (Ensemble) is a compromise in terms of cost at learning time while maintaining very good e ciency at classi cation time. The additional cost in learning is repaid by the possibility to create online learners, thus allowing incremental learning steps and only requiring to process the entire available training set once, i.e., the rst time. 4
Over the past few years, hate campaigns against individuals or groups, often minorities, have occurred on a variety of online platforms. Given the even extreme consequences that these incidents may cause to the victims, the scienti c community has recently spent increasing to realize e ective automated techniques able to recognize the presence of aggressive and hateful content within texts published online. To give the avor of how challenging this task is, the report in [18] summarises the ndings of the `Shared Task on Aggression Identi cation', a challenge organised in 2018 as part of a popular international conference on Computational Linguistics (COLING 2018). The task was to develop a classi er that could discriminate between three hate classes: Overtly Aggressive, Covertly Aggressive, and Non-aggressive texts. The organisers provided the participants with a dataset of 15,000 annotated comments from Facebook (in English and Hindi). Over a total of 130 participating teams, the best classi er obtained a weighted F-score of 0.64 for both Hindi and English.
Interestingly, the research team in [35, 36], in a parallel and independent way by the authors of this paper, considers a detection system combining convolutional and gated recurrent networks and tests it on a large collection of public Twitter datasets (in English), obtaining micro F1 scores that, for most of the datasets, improve the state-of-art performances (values range from 0.83 to 0.94). The considered architecture is di erent from that presented in Figure 1c. In fact, work in [35, 36] considers a cascade of one CNN layer and one GRU layer, without the ensemble layer here considered.
The authors of [25] propose an ensemble of Recurrent Neural Network classi ers, which incorporates not only features associated with the document, but also user-related information, such as a tendency towards racism or sexism. The approach has been evaluated on a publicly available corpus of 16k tweets in English, manually labelled as containing sexist, racist, or neutral messages [33]. The RNN classi ers ensemble achieves a weighted average F score equal to 0.93, succeeding in distinguishing racism and sexism messages from normal text.
Still focusing on Twitter, work in [
There exist some work which focuses on less popular platforms, such as the discussion board 4chan13, whose sub-board `Politically Incorrect' has often been linked to the alt-right movement. The authors of [13] rely on standard NLP tools and polarity-annotated lexicons to identify hateful content within the posts of that sub-board. The results of the analysis show an extraordinary level of expressed hatred, particularly for ethnic reasons.
In August, 2014, a harassment campaign, primarily conducted over
Twitter with the use of the hashtag #GamerGate, targeted several women in the
video game industry, mainly expressing sexist and anti-progressivism opinions.
By means of a methodology similar to that followed in [13], work in [
Still regarding harassment campaigns, a raising and worrying phenomenon, known as raiding, consists of organize and coordinate ad-hoc mobs aimed at disrupt a speci c social platform. The authors of [22] employ machine learning techniques to predict those YouTube videos which are likely to be raided by users of third-party hateful communities.
Focusing on the Italian language, [8] designed and developed the rst hate speech classi er for Italian, testing two di erent state-of-art approaches for sentiment analysis tasks (SVM and LSTM algorithms). Initially, the Facebook dataset introduced in Section 2.1 was annotated according to three distinct classes: comments could either be tagged as Strong Hate, Weak Hate, No Hate. Unfortunately, both SVM and LSTM were not able to discriminate well among the three classes. This was particularly true for the Strong Hate class. These results were probably due to the small number of Strong Hate documents in the dataset and the low level of annotators' agreement. To date, as also highlighted by [26, 28], dealing with multiple hate classes lead to a low agreement among the anno13 http://www.4chan.org/ tators, and thus, to scarce results in terms of automated classi cation. Instead, when considering two classes (Hate/No Hate) and only the documents for which at least 70% of the annotators were in agreement, both SVM and LSTM perform better, with a F1 score equal to 0.72 for the Hate class. 5
The growing spread of online hate campaigns against individuals and minorities quests for e ective techniques to automatically detect hate content. Parallel to standard text classi cation methodologies, mainly based on processing texts through conventional machine learning tools, we assist to the ourishing of attempts employing neural networks. In this work, we proposed six di erent classi cation con gurations and we evaluated their performances over datasets consisting of Italian Facebook posts and tweets. Results showed that those congurations where deep learning is used perform better than the others. This outcome is promising, since deep learning features the noticeable advantages to be independent from i) data re-training, ii) NLP machinery specialised for speci c languages, iii) the time-consuming feature engineering phase.
Given the encouraging results, as future work we aim at testing other DL con gurations, possibly over larger datasets. Furthermore, in order to assess the strength of our choices wrt related work, we will test the proposals available in the literature, including ours, on the same dataset. As an example, aiming at maintaining the dataset considered in this work, a promising starting point is to work on the architectures, and associated results, of the Hate Speech detection task at EVALITA, which took place in late December, 2018. tools for the arms race. In Proceedings of the 26th International Conference on World Wide Web Companion, WWW '17 Companion, pages 963{972, 2017. 7. Thomas Davidson, Dana Warmsley, Michael W. Macy, and Ingmar Weber. Automated hate speech detection and the problem of o ensive language. In Proceedings of the Eleventh International Conference on Web and Social Media, ICWSM 2017, Montreal, Quebec, Canada, May 15-18, 2017., pages 512{515, 2017. 8. Fabio Del Vigna, Andrea Cimino, Felice Dell'Orletta, Marinella Petrocchi, and Maurizio Tesconi. Hate me, hate me not: Hate speech detection on Facebook. In Proceedings of the First Italian Conference on Cybersecurity (ITASEC17), Venice, Italy, January 17-20, 2017., pages 86{95, 2017. 9. Emilio Ferrara, Onur Varol, Clayton Davis, Filippo Menczer, and Alessandro Flammini. The rise of social bots. Commun. ACM, 59(7):96{104, June 2016. 10. George Forman. An extensive empirical study of feature selection metrics for text classi cation. Journal of machine learning research, 3(Mar):1289{1305, 2003. 11. Jo~ao Gama, Indre_ Zliobaite_, Albert Bifet, Mykola Pechenizkiy, and Abdelhamid Bouchachia. A survey on concept drift adaptation. ACM Comput. Surv., 46(4):44:1{44:37, March 2014. 12. Yoav Goldberg. Neural Network Methods in Natural Language Processing. Morgan & Claypool Publishers, 2017. 13. Gabriel Emile Hine, Jeremiah Onaolapo, Emiliano De Cristofaro, Nicolas Kourtellis, Ilias Leontiadis, Riginos Samaras, Gianluca Stringhini, and Jeremy Blackburn. Kek, Cucks, and God Emperor Trump: A measurement study of 4chan's politically incorrect forum and its e ects on the web. In Eleventh International Conference on Web and Social Media, ICWSM 2017, Montreal, Quebec, Canada, May 15-18, 2017., pages 92{101, 2017. 14. Thorsten Joachims. Making large-scale support vector machine learning practical.
In Bernhard Scholkopf, Christopher J. C. Burges, and Alexander J. Smola, editors, Advances in Kernel Methods, pages 169{184. MIT Press, Cambridge, MA, USA, 1999. 15. Thorsten Joachims. Transductive inference for text classi cation using support vector machines. In ICML, volume 99, pages 200{209, 1999. 16. Rafal Jozefowicz, Oriol Vinyals, Mike Schuster, Noam Shazeer, and Yonghui Wu.
Exploring the limits of language modeling. arXiv preprint arXiv:1602.02410, 2016. 17. Yoon Kim. Convolutional neural networks for sentence classi cation. arXiv preprint arXiv:1408.5882, 2014. 18. Ritesh Kumar, Atul Kr. Ojha, Shervin Malmasi, and Marcos Zampieri. Benchmarking aggression identi cation in social media. In Proceedings of the First Workshop on Trolling, Aggression and Cyberbullying (TRAC-2018), pages 1{11. Association for Computational Linguistics, 2018. 19. La Repubblica (in Italian). Hater augura la morte ai gli di Bonucci: Forse un giorno vivremo in un mondo migliore, risponde il calciatore. https: //www.repubblica.it/sport/2018/09/13/news/hater\_augura\_la\_morte\ _ai\_figli\_di\_bonucci\_la\_risposta\_del\_calciatore-206310289/, September 13, 2018. 20. La Repubblica (in Italian). Insulti sui social a Chiara Bordi, la 18enne disabile candidata a Miss Italia. https://www.repubblica.it/spettacoli/people/ 2018/09/15/news/insulti_sui_social_a_chiara_bordi_la_18enne_disabile_ candidata_a_miss_italia-206553959/, September 18, 2018. 21. Daily Mail. Zuckerberg in Germany: No place for hate speech on Facebook. http://www.dailymail.co.uk/wires/ap/ article-3465562/Zuckerberg-no-place-hate-speech-Facebook.html, February, 26 2016. 22. Enrico Mariconti, Guillermo Suarez-Tangil, Jeremy Blackburn, Emiliano De Cristofaro, Nicolas Kourtellis, Ilias Leontiadis, Jordi Luque Serrano, and Gianluca Stringhini. You know what to do: Proactive detection of YouTube videos targeted by coordinated hate attacks. CoRR, abs/1805.08168, 2018. 23. Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Je Dean. Distributed representations of words and phrases and their compositionality. In C. J. C. Burges, L. Bottou, M. Welling, Z. Ghahramani, and K. Q. Weinberger, editors, Advances in Neural Information Processing Systems 26, pages 3111{3119.
Curran Associates, Inc., 2013. 24. Chikashi Nobata, Joel Tetreault, Achint Thomas, Yashar Mehdad, and Yi Chang.
Abusive language detection in online user content. In Proceedings of the 25th International Conference on World Wide Web, WWW '16, pages 145{153, 2016. 25. Georgios K. Pitsilis, Heri Ramampiaro, and Helge Langseth. Detecting o ensive language in tweets using deep learning. CoRR, abs/1801.04433, 2018. 26. Fabio Poletto, Marco Stranisci, Manuela Sanguinetti, Viviana Patti, and Cristina Bosco. Hate speech annotation: Analysis of an Italian Twitter corpus. In CLiC-it, 2017. 27. Jacob Ratkiewicz, Michael Conover, Mark Meiss, Bruno Goncalves, Alessandro Flammini, and Filippo Menczer. Detecting and tracking political abuse in social media. In Proc. 5th International AAAI Conference on Weblogs and Social Media (ICWSM), 2011. 28. Manuela Sanguinetti, Fabio Poletto, Cristina Bosco, Viviana Patti, and Marco Stranisci. An Italian Twitter corpus of hate speech against immigrants. In LREC, 2018. 29. Anna Schmidt and Michael Wiegand. A survey on hate speech detection using
Natural Language Processing. In SocialNLP@EACL, 2017. 30. Fabrizio Sebastiani. Machine learning in automated text categorization. ACM
Comput. Surv., 34(1):1{47, March 2002. 31. The Guardian. Facebook not protecting content moderators from mental trauma: lawsuit. https://www.theguardian.com/technology/2018/sep/24/ facebook-moderators-mental-trauma-lawsuit, September 24, 2018. 32. Simon Tong and Daphne Koller. Support vector machine active learning with applications to text classi cation. Journal of machine learning research, 2(Nov):45{ 66, 2001. 33. Zeerak Waseem and Dirk Hovy. Hateful symbols or hateful people? predictive features for hate speech detection on Twitter. In North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 88{93. Association for Computational Linguistics, 2016. 34. Wenpeng Yin, Katharina Kann, Mo Yu, and Hinrich Schutze. Comparative study of CNN and RNN for Natural Language Processing. arXiv preprint arXiv:1702.01923, 2017. 35. Ziqi Zhang and Lei Luo. Hate speech detection: A solved problem? The challenging case of long tail on Twitter. CoRR, abs/1803.03662, 2018. 36. Ziqi Zhang, David Robinson, and Jonathan Tepper. Detecting hate speech on Twitter using a convolution-GRU based deep neural network. In Aldo Gangemi, Roberto Navigli, Maria-Esther Vidal, Pascal Hitzler, Raphael Troncy, Laura Hollink, Anna Tordai, and Mehwish Alam, editors, The Semantic Web, pages 745{ 760, Cham, 2018. Springer International Publishing.