=Paper=
{{Paper
|id=Vol-2263/paper038
|storemode=property
|title=HanSEL: Italian Hate Speech Detection Through Ensemble Learning and Deep Neural Networks
|pdfUrl=https://ceur-ws.org/Vol-2263/paper038.pdf
|volume=Vol-2263
|authors=Marco Polignano,Pierpaolo Basile
|dblpUrl=https://dblp.org/rec/conf/evalita/PolignanoB18
}}
==HanSEL: Italian Hate Speech Detection Through Ensemble Learning and Deep Neural Networks==
https://ceur-ws.org/Vol-2263/paper038.pdf
HanSEL: Italian Hate Speech detection through
Ensemble Learning and Deep Neural Networks
Marco Polignano Pierpaolo Basile
University of Bari Aldo Moro University of Bari Aldo Moro
Dept. Computer Science Dept. Computer Science
via E. Orabona 4, 70125 Bari, Italy via E. Orabona 4, 70125 Bari, Italy
marco.polignano@uniba.it pierpaolo.basile@uniba.it
Abstract maggioranza: Support Vector Machine
(Hearst et al., 1998) (SVM con kernel
English. The detection of hate speeches, RBF), Random Forest (Breiman, 2001),
over social media and online forums, is Deep Multilayer Perceptron (Kolmogorov,
a relevant task for the research area of 1992) (MLP). Ogni classificatore è stato
natural language processing. This interest configurato utilizzando una strategia
is motivated by the complexity of the task greedy di ottimizzazione degli iper-
and the social impact of its use in real parametri considerando il valore di
scenarios. The task solution proposed in ”F1” calcolato su una suddivisione
this work is based on an ensemble of three casuale in 5-fold del set di training.
classification strategies, mediated by a Ogni frase è stata pre-elaborata affinchè
majority vote algorithm: Support Vector fosse trasformarta in formato word em-
Machine (Hearst et al., 1998) (SVM with beddings e TF-IDF. I risultati ottenuti
RBF kernel), Random Forest (Breiman, tramite cross-validation sul training set
2001), Deep Multilayer Perceptron (Kol- hanno mostrato un valore F1 pari a
mogorov, 1992) (MLP). Each classifier 0.8034 per le frasi estratte da Facebook
has been tuned using a greedy strategy of e 0.7102 per quelle di Twitter. Il codice
hyper-parameters optimization over the sorgente del sistema proposto può essere
”F1” score calculated on a 5-fold random scaricato tramite GitHub: https:
subdivision of the training set. Each sen- //github.com/marcopoli/
tence has been pre-processed to transform haspeede_hate_detect
it into word embeddings and TF-IDF bag
of words. The results obtained on the
cross-validation over the training sets have 1 Introduction and background
shown an F1 value of 0.8034 for Facebook
sentences and 0.7102 for Twitter. The In the current digital era, characterized by the large
code of the system proposed can be use of the Internet, it is common to interact with
downloaded from GitHub: https: others through chats, forums, and social networks.
//github.com/marcopoli/ Common is also to express opinions on public
haspeede_hate_detect pages and online squares. These places of discus-
sion are frequently transformed into ”fight clubs”
Italiano. L’individuazione di discorsi where people use insults and strong words in or-
di incitamento all’odio sui social media der to support their ideas. The unknown identity
e sui forum on-line è una sfida rile- of the writer is used as an excuse to fell free of
vante per l’area di ricerca riguardante consequences derived by attacking people only for
l’elaborazione del linguaggio naturale. their gender, race or sexual inclinations. A gen-
Tale interesse è motivato della complessità eral absence of automatic moderation of contents
del processo e dell’impatto sociale del can cause the diffusion of this phenomenon. In
suo utilizzo in scenari reali. La soluzione particular, consequences on the final user could be
proposta in questo lavoro si basa su un psychological problems such as depression, rela-
insieme di tre strategie di classificazione tional disorders and in the most critical situations
mediate da un algoritmo di votazione per also suicidal tendencies.
A recent survey of state of the art approaches that indicates the presence and absence of hate
for hate speech detection is provided by (Schmidt speeches. The task is organized into three sub-
and Wiegand, 2017). The most common systems tasks, based on the dataset used for training and
of speech detection are based on algorithms of text testing the participants’ systems:
classification that use a representation of contents
based on ”surface features” such as them available • Task 1: HaSpeeDe-FB, where only the
in a bag of words (BOW) (Chen et al., 2012; Xu et Facebook dataset can be used to classify the
al., 2012; Warner and Hirschberg, 2012; Sood et Facebook test set
al., 2012). A solution based on BOW is efficient
• Task 2: HaSpeeDe-TW, where only the
and accurate especially when n-grams have been
Twitter dataset can be used to classify the
extended with semantic aspects derived by the
Twitter test set
analysis of the text. (Chen et al., 2012) describe an
increase of the classification performances when • Task 3: Cross-HaSpeeDe, which can be fur-
features such as the number of URLs, punctua- ther subdivided into two sub-tasks:
tions and not English words are added to the vec-
torial representation of the sentence. (Van Hee et 1. Task 3.1: Cross-HaSpeeDe FB, where
al., 2015) proposed, instead, to add as a feature only the Facebook dataset can be used
the number of positive, negative and neutral words to classify the Twitter test set
found in the sentence. This idea demonstrated 2. Task 3.2: Cross-HaSpeeDe TW,
that the polarity of sentences positively supports where only the Twitter dataset can be
the classification task. These approaches suffer used to classify the Facebook test set
from the lack of generalization of words contained
into the bag of words, especially when it is cre- The Facebook and Twitter datasets released for
ated through a limited training set. In particular, the task consist of a total amount of 4,000 com-
terms found in the test sentences are often missing ments/tweets, randomly split into development
in the bag. More recent works have proposed word and test set, of 3,000 and 1,000 messages respec-
embeddings (Le and Mikolov, 2014) as a possi- tively. Data are encoded in a UTF-8 with three
ble distributional representation able to overcome tab-separated columns, each one representing the
this problem. This representation has the advan- sentence id, the text and the class (Fig. 1).
tage to transform semantically similar words into
id text hs
a similar numerical vector. Word embeddings are
8 Io votero NO NO E NO 0
consequently used by classification strategies such
36 Matteo serve un colpo di stato. 1
as Support Vector Machine and recently by deep
learning approaches such as deep recurrent neural Table 1: Examples of annotated sentences.
networks (Mehdad and Tetreault, 2016). The so-
lution proposed in this work reuse the findings of
(Chen et al., 2012; Mehdad and Tetreault, 2016) 3 Description of the system
for creating an ensemble of classifiers, including
a deep neural network, which works with a com- The system proposed in this work is HanSEL:
bined representation of word embeddings and a a system of Hate Speech detection through En-
bag of words. semble Learning and Deep Neural Networks. We
decided to approach the problem using a classic
2 Task and datasets description natural language processing pipeline with a final
task of sentences classification into two exclusive
The hate speech detection strategy proposed in classes: hate speech and not hate speech. The data
HAnSEL has been developed for HaSpeeDe (Hate provided by task organizers are obtained crawling
Speech Detection) task organized within Evalita social network, in particular, Facebook and Twit-
2018 (Caselli et al., 2018), which is going to ter. The analysis of the two data sources showed
be held in Turin, Italy, on December 12th-13th, many possible difficulties to face in the case of us-
2018 (Bosco et al., 2018). HaSpeeDe consists in ing approaches based on Italian lexicons of hate
the annotation of messages from social networks speeches. In particular, we identified the follow-
(Twitter and Facebook) with a boolean label (0;1) ing issues:
• Repeated characters: many words in- We follow the same step of pre-processing ap-
cludes characters repeated many times for plied by Tripodi (Tripodi and Li Pira, 2017) to
emphasizing the semantic meaning of the transform the sentence of the task datasets into
word. As an example, the words ”nooooo”, word embeddings. In particular, we applied the
”Grandeeeee”, ”ccanaleeeeeeeeeeeeeeee” following Natural Language Processing pipeline:
are found in the training of Facebook
• Reduction of repeated characters: we scan
messages.
each sentence of the datasets (both training
• Emoji: sentences are often characterized by and test). For each sentence, we obtain words
emoji such as hearts and smiley faces that are merely splitting it by space. Each word is an-
often missing in external lexicons. alyzed, and characters repeated three times or
more are reduced to only two symbols, try-
• Presence of links, hashtags and mentions:
ing to keep intact word that naturally includes
this particular elements are typical of the
doubles.
social network language and can introduce
noise in the data processing task. • Data cleansing: we transformed the words is
lowercase and following we removed from
• Length of the sentences: many sentences are
each sentences links, hashtags, entities, and
composed by only one word or in general,
emoji
they are very short. Consequently, they are
not expressive of any semantic meaning. The normalized sentences are consequently to-
The complexity of the writing style used in hate kenized using the TweetTokenizer of the NLTK
speech sentences guided us through the idea library 1 . For each sentence we averaged the
to do not use an approach based on standard word2vec vectors correspondent of each token, re-
lexicons and to prefer supervised learning strate- moving during each sum the centroid of the whole
gies on the dataset provided by the task organizers. distributional space. This technique is used for
mitigating the problems of loss of information due
Sentence processing. to the operation of averaging the semantic vectors.
We decide to represent each sentence as a con- The two bags of words (Facebook and Twitter)
catenation of a 500 features word embedding are, instead, created directly on the sentences
vector and a 7,349 size bag of words for Facebook without any pre-processing, also if during the
messages and 24,866 size bag of words for Twitter tuning of the architecture we had tried some con-
messages. In particular, the word-embedding pro- figurations that include bag of words without stop
cedure used is word2vec introduced by Mikolov words, with lowercase letters and processed by
(Mikolov et al., 2013). This model learns a Snowball stemmer algorithm 2 without obtaining
vector representation for each word using a neural breaking results. The n-gram size considered for
network language model and can be trained the construction of the bag is in the range of 1 to
efficiently on billions of words. Word2vec allows 3. The final representation of each sentence of the
being a very efficient data representation in text dataset is consequently obtained concatenating
classification due to its capability to create very the word2vec vector and the correspondent bag
similar vectors for words strongly semantically of words. Sentences too shorts that cannot be
related. The Italian word embeddings used in transformed into word2vec as a consequence
this work are provided by Tripodi (Tripodi and of the absence of all the tokens of the sentence
Li Pira, 2017). The author trained the model on a have been classified using only the bag of words
dump of the Italian Wikipedia (dated 2017.05.01), representation.
from which only the body text of each article is
used. The corpus consists of 994,949 sentences Classification strategy.
that result in 470,400,914 tokens. The strategy of HAnSEL is based on a classification process that
the creation of word embeddings is CBOW with uses three different classification strategies medi-
the size of the vectors equal to 500, the window ated by a hard majority vote algorithm. A stack-
size of the words contexts set to 5, the minimum ing of classifiers with a Random Forest blender
number word occurrences equal to 5 and the 1
https://www.nltk.org/data.html
2
number of negative samples set to 10. http://snowball.tartarus.org/texts/quickintro.html
has also been considered during the design of according to their feature values. In order to
HAnSEL architecture but, the internal evaluation classify an unseen item, the three is navigated
runs on 5-fold cross-validation of the training set until reaching the leaf and then the ration of
showed us low performances of the approach. training items of class k in that leaf is used as
This analysis is not detailed more in this work due a class probability.
to the page limitations of it. In order to design the
ensemble, we analyzed the performances of some • Random forest classifier (RF). It is an en-
of the most popular classification algorithms for semble of Decision Trees trained on different
the text categorization task. In particular, we con- batches of the dataset that uses averaging to
sidered: improve the predictive accuracy and control
over-fitting. A typical parameter is the num-
• Logistic regression with stochastic gradient ber of threes to use in order to balance the
descendent training (SGD). It has the advan- precision of the algorithm and the random-
tage to be very efficient with large datasets ness to obtain a good level of generalization
considering, during the training, one instance of the model.
per time independent by others. It uses the
• Multi-layer Perceptron classifier (MLP). This
gradient descendent as optimization function
model is a classical architecture of a deep
for learning the optimal weight of the sepa-
neural network. It is composed by one layer
ration function of the distributional space of
of inputs, one layer of linear threshold units
items. In literature, it has been successfully
(LTU) as output and many hidden layers of an
used for tasks of text classification, especially
arbitrary number of LTU plus one bias neu-
with binary classification problems.
ron fully connected each other. The weights
• C-Support Vector Classification (SVC). It is learned by each neuron (perceptron) are up-
the standard Support Vector Machine algo- dated through back-propagation using a gra-
rithm applied for the classification task. It is dient descendent strategy. Important parame-
a powerful approach that supports linear and ters to configuring are the number of hidden
non-linear classification function. Moreover, layers, the number of training epochs and the
through the C parameter, it is possible to de- L2 penalty (regularization term) parameter.
cide how much the margin of classification We evaluated the performance of the algorithms
could be significant and consequently sensi- just described using a default configuration and a
tive to outliers. The implementation is based 5-fold cross validation over the Facebook training
on libsvm, and we evaluated different config- set. Moreover, we set the random seed equal to 42
urations of the algorithm: polynomial func- for obtaining at each run always the same folder
tion with 2 and 3 degree, RBF kernel and dif- subdivision.
ferent values of the C parameters. Tab. 3 shows the results obtained by the dif-
• K-nearest neighbors vote (KNN). This clas- ferent classification algorithms during their pre-
sic and versatile algorithm is based on the liminary analysis considering the macro F1 score
concept of similarity among items accord- as in the task specifications. The values obtained
ing to a distance metric. In particular, for do not point out significant statistical differences
an unseen item, the k most similar items of among the approaches, but we decided to investi-
the training set are retrieved, and the class, gate more the top three scored algorithms: SVM
provided as output, is obtained by the major- with an RBF kernel, Random Forest with 300
ity vote of the neighborhoods. Despite the trees, MLP with 2,000 hidden layers. In general,
simplicity of the algorithm it is often used in we observed that linear algorithms obtain a high
tasks of text classification. score for the task supporting our idea that linguis-
tic features are enough for defining a clear sepa-
• A decision tree classifier (DT). This approach ration among the sentences of hate and not hate
is another popular strategy of classification speeches. In order to identify an optimal config-
used especially when it is required to visu- uration of the algorithms, we trained our models
alize the model. The DT algorithm works using a greedy search approach. For each algo-
splitting items into a different path of the tree rithm, we performed 100 training runs with pa-
Not HS HS
Precision Recall F1-score Precision Recall F1-score Macro F1 Pos.
Task 1 0,6981 0,6873 0,6926 0,8519 0,8581 0,855 0,7738 7
Task 2 0,7541 0,8801 0,8122 0,6161 0,4012 0,4859 0,6491 14
Task 3.1 0,7835 0,2677 0,3991 0,3563 0,8456 0,5013 0,4502 11
Task 3.2 0,3674 0,8235 0,5081 0,7934 0,3234 0,4596 0,4838 8
Table 2: Final scores obtained during the HASpeeDe challenge
Algorithm Macro F1 score max iter=184,solver=’adam’,
LR 0.780444109 warm start=False)
SVC-rbf - C= 1 0.789384136
SVC-poly 2 C=1 0.758599844 The models are consequently used in a vot-
SVC-poly 3 C=1 0.667374386 ing classifier configured for using a hard majority
KNN - 3 0.705064332 vote algorithm. The ensemble obtains an F1 value
KNN - 5 0.703990867 of 0.8034 for Facebook sentences and 0.7102 for
KNN - 10 0.687719117 Twitter using the 5-fold subdivision of the training
KNN - 20 0.663451598 sets. The implementation of the system has been
DT 0.68099986 realized into Python language and using the scikit-
RF-50 0.75219596 learn 0.20 machine learning library 3 .
RF-100 0.764247578
RF-300 0.787778421 4 Results and discussion
RF-500 0.768494151 HanSEL has been used for classifying the data
MLP-1000 0.766835616 provided as a test set for each of the three special-
MLP-2000 0.791230474 ized tasks of HaSpeeDe competition. Tab. 2 shows
MLP-3000 0.76952709 the final results obtained by our system in the chal-
Table 3: Classification algorithms on Facebook lenge. It is possible to observe that the system
training set using 5-fold cross validation. well performed for Task 1 and Task 3.2 which in-
volve the classification of Facebook messages. In
particular, it emerges that HanSEL performs bet-
rameters randomly selected from a range of values ter fot hate speeches sentences than for not hate
preliminary defined. Each run has been evaluated, speeches probably a consequence of the presence
considering the macro F1 score, on the training of many clear hate words used in this type of mes-
set using the same strategy of cross-validation al- sages such as ”sfigati” and ”bugiardo” in that cate-
ready described before. At the end of the 100 runs gory of textual sentences. A symmetrical situation
the model that achieve the best results has been is obtained for Task 2 and Task 3.2 that involves
stored and later used in our final ensemble of clas- Twitter messages. In this scenario, the significant
sifiers. The final configurations obtained for the use of specific hashtags, irony, and entities instead
three strategies are the following: of clear hate words has made difficult the identi-
fication of hate speeches. The cross-classification
• SVC(C=1.76800710431488, task has, moreover, stressed the generalization of
gamma=0.1949764030136127, kernel=’rbf’) the system. It has been observed that the writ-
ing style of the two social networks strongly in-
• RandomForestClassifier(bootstrap=False, fluences the classification performance, especially
max depth=30,max features=’sqrt’, when the models are trained on a small training
min samples leaf=2,min samples split=2, set, as in our case. Finally, the optimization of
n estimators=200, warm start=False) the models inside the ensemble has been stressed
more on the Facebook dataset consequently over-
• MLPClassifier(alpha=0.5521952082781035,
fitting on the characteristics of that type of mes-
early stopping=False,
sages. The outcomes achieved for the challenge
hidden layer sizes=2220,
3
learning rate init=0.001, http://scikit-learn.org/stable/
allow us to deduce important consideration for fur- Security, Risk and Trust (PASSAT), 2012 Interna-
ther developments of the system. In particular, we tional Conference on and 2012 International Con-
fernece on Social Computing (SocialCom), pages
consider essential to mix the two datasets in or-
71–80. IEEE.
der to allow the models to generalize better con-
sidering the two different sources of data. More- Marti A. Hearst, Susan T Dumais, Edgar Osuna, John
over, extra features regarding hashtags, entities, Platt, and Bernhard Scholkopf. 1998. Support vec-
tor machines. IEEE Intelligent Systems and their ap-
and links can be helpful for obtaining better results plications, 13(4):18–28.
with Twitter messages.
Věra Kolmogorov. 1992. Kolmogorov’s theorem
5 Conclusion and multilayer neural networks. Neural networks,
5(3):501–506.
The HaSpeeDe competition has been a perfect sce-
Quoc Le and Tomas Mikolov. 2014. Distributed rep-
nario for developing and testing solutions for the
resentations of sentences and documents. In Inter-
social problem of hate speeches on social media national Conference on Machine Learning, pages
and, in particular, for them in the Italian language. 1188–1196.
In our work, we presented HAnSEL a system based
Yashar Mehdad and Joel Tetreault. 2016. Do charac-
on an ensemble of classifiers that includes the Sup- ters abuse more than words? In Proceedings of the
port Vector Machine algorithm, Random Forests, 17th Annual Meeting of the Special Interest Group
and a Multilayers Perceptron Deep Neural Net- on Discourse and Dialogue, pages 299–303.
work. We formalize messages as a concatenation
Tomas Mikolov, Kai Chen, Greg Corrado, and Jef-
of word2vec sentence vectors and a TF-IDF bag of frey Dean. 2013. Efficient estimation of word
words. Results showed the efficacy of the solution representations in vector space. arXiv preprint
in a scenario that uses clear offensive words such arXiv:1301.3781.
as Facebook messages. On the contrary, there is Anna Schmidt and Michael Wiegand. 2017. A survey
a large margin of improvements for the classifica- on hate speech detection using natural language pro-
tion of Tweets. The future direction of the work cessing. In Proceedings of the Fifth International
will surely investigate the use of more data and se- Workshop on Natural Language Processing for So-
cial Media, pages 1–10.
mantic features for allowing classification meth-
ods to create a more general model. Sara Sood, Judd Antin, and Elizabeth Churchill. 2012.
Profanity use in online communities. In Proceed-
ings of the SIGCHI Conference on Human Factors
References in Computing Systems, pages 1481–1490. ACM.
Cristina Bosco, Felice Dell’Orletta, Fabio Poletto, Rocco Tripodi and Stefano Li Pira. 2017. Anal-
Manuela Sanguinetti, and Maurizio Tesconi. 2018. ysis of italian word embeddings. arXiv preprint
Overview of the EVALITA 2018 HaSpeeDe Hate arXiv:1707.08783.
Speech Detection (HaSpeeDe) Task. In Tom-
maso Caselli, Nicole Novielli, Viviana Patti, and Cynthia Van Hee, Els Lefever, Ben Verhoeven, Julie
Paolo Rosso, editors, Proceedings of the 6th evalua- Mennes, Bart Desmet, Guy De Pauw, Walter Daele-
tion campaign of Natural Language Processing and mans, and Véronique Hoste. 2015. Detection and
Speech tools for Italian (EVALITA’18), Turin, Italy. fine-grained classification of cyberbullying events.
CEUR.org. In International Conference Recent Advances in
Natural Language Processing (RANLP), pages 672–
Leo Breiman. 2001. Random forests. Machine learn- 680.
ing, 45(1):5–32.
William Warner and Julia Hirschberg. 2012. Detect-
Tommaso Caselli, Nicole Novielli, Viviana Patti, and ing hate speech on the world wide web. In Proceed-
Paolo Rosso. 2018. EVALITA 2018: Overview of ings of the Second Workshop on Language in Social
the 6th Evaluation Campaign of Natural Language Media, pages 19–26. Association for Computational
Processing and Speech Tools for Italian. In Tom- Linguistics.
maso Caselli, Nicole Novielli, Viviana Patti, and
Paolo Rosso, editors, Proceedings of Sixth Evalua- Jun-Ming Xu, Kwang-Sung Jun, Xiaojin Zhu, and
tion Campaign of Natural Language Processing and Amy Bellmore. 2012. Learning from bullying
Speech Tools for Italian. Final Workshop (EVALITA traces in social media. In Proceedings of the 2012
2018), Turin, Italy. CEUR.org. conference of the North American chapter of the as-
sociation for computational linguistics: Human lan-
Ying Chen, Yilu Zhou, Sencun Zhu, and Heng Xu. guage technologies, pages 656–666. Association for
2012. Detecting offensive language in social me- Computational Linguistics.
dia to protect adolescent online safety. In Privacy,