SOCIAL NETWORK HATE SPEECH DETECTION FOR AMHARIC LANGUAGE

Dhinaharan Nagamalai et al. (Eds) : NATL, CSEA, DMDBS, Fuzzy, ITCON, NSEC, COMIT - 2018
pp. 41–55, 2018. © CS & IT-CSCP 2018 DOI : 10.5121/csit.2018.80604
SOCIAL NETWORK HATE SPEECH
DETECTION FOR AMHARIC LANGUAGE
Zewdie Mossie1
and Jenq-Haur Wang2
1
Department of International Graduate Program in Electrical Engineering and
Computer Science,
National Taipei University of Technology, Taipei, Taiwan
2
Department of Computer Science and Information Engineering,
National Taipei University of Technology, Taipei, Taiwan
ABSTRACT
The anonymity of social networks makes it attractive for hate speech to mask their criminal
activities online posing a challenge to the world and in particular Ethiopia. With this ever-
increasing volume of social media data, hate speech identification becomes a challenge in
aggravating conflict between citizens of nations. The high rate of production, has become
difficult to collect, store and analyze such big data using traditional detection methods. This
paper proposed the application of apache spark in hate speech detection to reduce the
challenges. Authors developed an apache spark based model to classify Amharic Facebook
posts and comments into hate and not hate. Authors employed Random forest and Naïve Bayes
for learning and Word2Vec and TF-IDF for feature selection. Tested by 10-fold cross-
validation, the model based on word2vec embedding performed best with 79.83%accuracy. The
proposed method achieve a promising result with unique feature of spark for big data.
KEYWORDS
Amharic Hate speech detection, Social networks and spark, Amharic posts and comments
1. INTRODUCTION
A major bottleneck for promoting use of computers and the Internet is that many languages lack
the basic tools that would make it possible for people to access ICT in their own language. The
status of language processing tools for European languages[2] states that only English, French
and Spanish have sufficient basic tools. Thus the vast majority of the World’s languages are still
under-resourced in that they have few or no language processing tools and resources which
particularly true for sub Saharan African languages. However, the evolution of the Internet and of
social media texts, such as Twitter, YouTube and Facebook messages, has created many new
opportunities for creating such tools, but also many new challenges [1].Amharic is one of the sub-
Saharan countries Ethiopian’s working language which is written left-to-right in its own unique
script which lacks capitalization and in total has 275 characters mainly consonant-vowel pairs. It
is the second largest Semitic language in the world after Arabic and spoken by about 40% of the
population as a first or second language [3] but current population estimated to 102 million. In

42 Computer Science & Information Technology (CS & IT)
spite of its relatively large number of speakers, Amharic is still a language for which very few
computational linguistic resources have been developed, and very little has been done in terms of
making useful higher level Internet or computer based applications.
This paper focus only on hate speech detection from social media posts and comments. Recent
advances in mobile computing and the Internet have resulted in an increase in use of social media
to communicate, express opinions, interact with other, and to find and share information [4].
While social media provides an important avenue for communication to take place easily and
efficiently, it also acts as a means of spreading hate speech online. Inherent characteristics of the
Internet largely contribute to the misuse of social network to transmit and propagate hate speech.
Hate messages are prevalent and challenging in the Ethiopian online community as individuals
spread hate messages hiding behind their screens. The government of Ethiopia oversee and
monitor content in social network in a bid to govern hate speech through one time interruption of
the internet service. Research conducted by Amnesty International and the Open Observatory of
Network Interference (OONI) between June and October 2016 shows that access to WhatsApp
and others was blocked, as well as at least 16 news outlets [6]. It is an open secret that the recent
widespread hate speech and call for violence particularly targets persons of a particular group
[5].In this regard no work is done before and the first for Amharic language even though the work
of [36]done from the social science perspective. It is therefore, of critical importance to monitor
and identify instances of hate speech, as soon as possible to prevent their spread and possible
unfolding into acts of violence or hate crimes and destroys the lives of individuals, families,
communities and the country.
The proposed method used Word2Vec and TF-IDF for feature selection and Naïve Bayes and
Random forest machine learning algorithms known for hate speech detection performance. The
rest of this paper is organized as follows. Section 2 reviews related work on hate speech
detection. The method and data preprocessing steps are described in detail in Section 3.
Architectural design and experimentations are illustrated and discussed in Section 4. Finally,
conclusion and future work in Section 5.
2. RELATED WORK
2.1 Hate Speech on Social Media
Online spaces are often exploited and misused to spread content that can be degrading, abusive,
or otherwise harmful to people. Hateful speech has become a major problem for every kind of
online platform where user-generated content appears from the comment sections of news
websites to real-time chat sessions. Legal and academic literature generally deﬁnes hate speech as
speech or any form of expression that expresses hatred against a person or group of people
because of a characteristic they share, or a group to which they belong [7]. But, there is no
consensus deﬁnition because of prevailing social norms, context, and individual and collective
interpretation. A recent study define hate speech as speech which either promotes acts of violence
or creates an environment of prejudice that may eventually result in actual violent acts against a
group of people[8]. In the case of Ethiopia the use of hateful words with an intention to bring
about hatred against a group of people based on their ethnicity, political attitude, religion and
socio -economic are prevailing [36].

Computer Science & Information Technology (CS & IT) 43
2.2 Social Media Definition of Hate Speech
• Hate speech is to incite violence or hate: The several definitions use slightly different
terms to describe when hate speech occurs. The majority of the definitions point out that
hate speech is to incite violence or hate towards a minority (Code of conduct, ILGA,
YouTube and Twitter)
• Hate speech is to attack or diminish: Additionally, some other definitions state that hate
speech is to use language that attacks or diminishes these groups (Facebook,YouTube,
and Twitter).
After consulting those papers, authors use these dimensions of analysis to define what is hate
speech in the scope of this paper.
2.3 Existing Techniques Used in Hate Speech Detection in Social Media
The study of hate speech detection has been growing only in the few last years. However, some
studies have already been conducted in few languages. Papers focusing algorithms for hate
speech detection, and also other studies focusing on related concepts, can give us insight about
which features to use in this classification task. Therefore, authors allocate this specific section to
describe the features already employed in previous works dividing into two categories: general
features used in text mining and specific hate speech detection features.
Dictionaries and lexicons: The majority of the papers authors found try to adapt strategies
already known in text mining to the specific problem of hate speech detection. The work
categorize the features as the features commonly used in text mining which is dictionaries and
lexicons. This approach consists in making a list of words that are searched and counted in the
text. In the case of hate speech detection this has been conducted using content words such as
insult and swear words, reaction words, and personal pronouns [24], number of disrespectful
words in the text, with a dictionary that consists of words for English language including
acronyms and abbreviations [26], label specific features which consisted in using frequently used
forms of verbal abuse as well as widely used stereotypical words[27], Ortony lexicon was also
used for negative affect detection ( list of words denoting a negative connotation and can be
useful because not every rude comment necessarily contains bad language and can be equally
harmful) [11].
Bag-of-words(BOW): Another model similar to dictionaries is the use of bag-of-words [9,10,
11]. In this case, a corpus is created based on the words that are in the training data, instead of a
pre-defined set of words, as in the dictionaries. The disadvantages of this kind of approaches is
that the word sequence is ignored, and also it’s syntactic and semantic content. Therefore,it can
lead to misclassification if the words are used in different contexts. To overcome this limitation n-
grams were implemented. N-grams are one of the most used techniques in hate speech automatic
detection and related tasks [11, 12, 13, 14, 15]. In a study character ngram features proved to be
more predictive than to kenn-gram features, for the specific problem of abusive language
detection [16].
TF-IDF was also used in this kind of classification problems. It is a measure of the importance of
a word in a document within a corpus and increases in proportion to the number of times that a

word appears in the document. However, it is distinct from a bag of words, or n-gram, because
the frequency of the term is off-setted by the frequency of the word in the corpus, which
compensates the fact that some words appear more frequently in general [17].
Part-of-speech (POS) approaches also make it possible to improve the importance of the context
and detect the role of the word in the context of a sentence. These approaches consist in detecting
the category of the word, for instance, personal pronoun (PRP), Verb non-third person singular
present form (VBP), Adjectives (JJ), Determiners (DT), Verb base forms (VB). Part of speech has
also been used in hate speech detection problem even though proved to cause confusion in the
class’s identification [14]. It was also used to detect sentences such as “send them home”, “get
them out” or “should be hung” [18].
Word Embedding: Deep learning techniques are recently being used in text classification and
sentiment analysis with high accuracy [28]. One of the approaches of this is word embedding
which allows finding both semantic and syntactic relation of words, which permits the capturing
of more refined attributes and contextual cues that are inherent in human language. Therefore,
Word2Vec [19], an unsupervised word embedding-based approach to detect semantic and
syntactic word relations was used. Word2Vec is a two-layer neural network that operates on a set
of texts to initially establish a vocabulary based on the words included in such set more times
than a user-defined threshold to eliminate noise. According to [19] 50-300 dimensions can model
hundreds of millions of words with high accuracy. Possible methods to build the actual model are
CBOW (i.e., Continuous bag of words), which uses context to predict a target word, and Skip-
gram, which uses a word to predict a target context. Skip-gram works well with small amounts of
training data and handles rare words or phrases well, while CBOW shows better accuracy for
frequent words and is faster to train. Word embedding combined with Convolutional Neural
Networks (CNN) show better performance [20, 28]. Authors [26] use a paragraph2vec approach
to classify language on user comments as abusive or clean and also to predict the central word in
the message. Alternatively, other authors propose comment embedding to solve this problem
[27]. FastText is also being used [28] in a problem that sentences must be classified and not
words. Sentiment Analysis bearing in mind that hate speech has a negative polarity, authors have
been putting the sentiment as a feature for hate speech detection [15, 23, 24, 25, 31,].
2.4 Algorithms Commonly Used For Hate Speech Detection
Consulting different sources on algorithms of hate speech detection are the focus of this section,
because authors aim to work in this specific topic. In the majority of the works the used language
is English. However, there were some researched works done for languages Dutch [21]and Italian
[22]to author’s knowledge. The most common approach found in the work of [15] as a literature
review consists in building a machine learning model for hate speech classification. It is found
that the most common algorithms used are SVM, Random Forests, Decision Trees, logistic
regression, Naïve Bayes and Deep learning respectively on the use of frequency by authors. The
data classification is based on general hate speech, racism, sexism, religion, anti-Semitism,
nationality, politics and socio-economics status respectively on the categorization use of
frequency. Authors propose Random Forest and Naïve Bayes for their good performance.

3. PROPOSED METHODOLOGY AND DATA COLLECTION
Aiming at classifying the hate level across Facebook for Amharic language users, authors have
built a corpus of comments retrieved from Facebook public pages of Ethiopian newspapers,
individual politicians, activist, TV and radio broadcast and groups. These pages typically posts
discussions spanning across a variety of political and religious topics. By doing so, authors could
capture both casual conversations and politically hated posts and comments. Authors have
employed a versatile Facebook crawler, which exploits the Graph API to retrieve the content of
the comments from Facebook posts using Facepager. Facebook is selected to collect data from
social media for the following reasons. Facebook is the most important platform for reaching out
to online audiences, and especially the youth. Comparative studies have shown how in countries
with limited Internet penetration, like Ethiopia, Facebook has become almost a synonym for the
Internet, a platform through which users access information, services, and participate in online
communications.
3.1 Data Preparation and Annotation
Authors then preprocessed the posts and comments according to the following rules:
• Only kept comments that were in Amharic and all punctuations were removed by passing
to the apache spark map function
• All null values are also removed with isNull attribute of apache spark DataFrame
• Checked to assure that no repetitions with the same text by passing to the map function
using distinct attribute available on apache RDD and DataFrame
• Removed the HTML and different symbols in the same way using apache spark since
authors focus only on texts
• All elongations were removed to the same fixed size character based on the nature of
Amharic language and finally Trim text as final step
After all the above preprocessing authors consider the following three bases for future annotation:
(1) Discourse analysis:-places the text in its wider political, ethnicity, socio-economic and
religious context in order to understand the currents of thought which illustrate and rationalize
why it is to be considered hateful or not.
(2) Content analysis:-analyses the text deemed not hate and hate in order to pick out the key
semantic components and targets of the speech. This can then be coded and quantitative
techniques applied to draw wider patterns and trends.
(3) Automated techniques: - a relatively novel method of tracking hate speech that can be
usefully employed to mine high volumes of text from different sources to search for keywords
which are highly indicators of hate speech in an efficient manner which authors followed in
labeling the collected data. After the initial cleaning authors got 25,890 posts and comments

available, however authors sampled to be 10, 000 due to the limitation in resources for the
annotation task.
3.2 Annotation Instructions
Despite the differences between the previous studies that analyzed in the related work, the
majority of the described works present instructions for the annotation task. Some authors point
out that having vague annotation guidelines [8, 30]is a problem for hate speech detection due to
the complexity of the task. In this work, authors prepared a complete set of annotation
instructions in chart in order to better standardize the annotation procedure and to make clear all
hate and not hate speech related category concepts. A set of instructions and examples that
contain the indicators of the category was defined in figure 1. These are based on the definitions,
rules and examples that presented already in the related work. The annotators were given the
instructions as guidelines in the classification of the messages
Figure 1: Hierarchical Structure of Dataset Annotation
In the work of [32] the training data was hand-coded and manually annotated and admits the
potential for hard-to-trace bias in the hate speech categorization. The study concerned the
detection of racism using a Naive Bayes classifier. The work established the definitional
challenge of hate speech by showing annotators could agree only 33% of the time on texts
purported to contain hate speech. Another considered the problem of detecting anti-Semitic
comments in Yahoo news groups using support vector machines [29].
In this work first, authors consider a definition of hateful speech that could be practically useful to
platform operators of social media and previous work definitions. Second, develop a general
method of hierarchical annotation method shown in figure 1 for selected annotators of 3 PHD,2
MSC students and 1 assistant professor from Amharic Language studies. The annotators were
instructed to use the chart originates from the figure 1.In addition to the annotation rules the Kapa
decision agreement based on the Cohen's kappaκ statistic which is an estimate of the population
coefficient between 0 ≤ κ ≤ 1 [32] is also used. This work show how the values are interpreted?
What does a specific kappaκ value mean?

Table 1: kappaκ values
Nominal Poor Slight Fair Moderate Substantial Almost perfect
Kappa vlaue 0.0 0.20 0.40 0.60 0.80 10.
Kappaκ agreement < 0 less than chance agreement, 0.01–0.20 Slight agreement, 0.21- 0.40 Fair
agreement, 0.41-0.60 Moderate agreement, 0.61-0.80 Substantial agreement, 0.81-0.99, almost
perfect agreement. Not everyone would agree which one is “good” agreement but as commonly
cited scale is kappa value of 0.57 is in the “moderate” agreement range for better agreement.
Remember that perfect agreement would equate to a kappa of 1, and chance agreement would
equate to 0. Given that the majority of comments has been annotated by more than one annotator,
authors have also computed the kappaκ inter-annotator agreement metric [33], which measures
the level of agreement of different annotators on a task. In this case, considering 1,821 comments
that received annotations from all the 6 annotators and obtain = 0.64 when discriminating over
two classes and the work of [26] using number of disrespectful words in the text, with a
dictionary that extracted from the annotated dataset and identified by the language experts. Then
the dataset becomes larger than before which is 6, 120 to be used for this work.
Table 2: samples of kappaκ inter-annotator agreement result
Language Amharic Comment Text and its English translation class
Amharic መፍትሄውጎረቤትህያለውንትግሬመግደልነው Hate
English The solution is to kill the neighboring Tigrian
Amharic ኦሮሞየአማራጠላትነው Hate
English Oromo is an enemy of Amhara
Amharic አንድአማራለሁሉምአማራ Nohate
English One Amhara to All Amhara
4. ARCHITECTURAL DESIGN AND EXPERIMENTATION
An Apache Spark Standalone cluster was used for data preparation and developing models for
machine learning classification which is suitable for big data processing like Facebook data.
Spark ML pipeline is used in providing a set of tokenization mechanisms. In addition, Spark
offers modules for feature selection and machine learning MLLIB library. Python programming
language was used for both preparation of dataset and machine learning with RDD and
DataFrame file format used as the back end for storing lazy operations which is ideal for large
data. Spark designed to efficiently store RDD data while providing powerful MAP, Reduce and
filter transformation operations and take actions for further process as shown in the figure 2.
The model was trained using 4,882 posts and used to correctly classify Facebook data according
to the two classes mentioned above prepared based on the requirements of Naïve Bayes and
Random forest algorithms. These classifier were selected based on previous work result in related
work for English and other languages.

Figure 2: Architectural Design of Amharic hate speech detection
Table 2: Dataset information for the paper (new)
Training Dataset
Nohate Hate Total
2,629 2,253 4,882
Test Dataset
667 571 1,238
3,296 2,824 6,120
4.1 Feature Selection
This involved selecting a subset of relevant features that would help in identifying hate and no
hate posts and can be used in the modeling of the classification problems. Authors use Word2Vec
as used in [11, 12, 13, 14] for such work and text classification [34]. TF-IDF [16] also used in
text classification by different authors for feature selection in other tools. But authors propose to
use both of them for Apache Spark feature selection and transformation API. The main feature of
interest for this work is comments and posts sentiment of users towards hate speech in social
media. The classification is supervised learning task because the objective is to use machine
learning to automatically classify comments/posts into categories based on previously labelled
comments and posts [11]. Author’s contribution is preparation of new dataset, using tf-idf and
word2vec as feature extraction, first in its kind, for the Amharic language hate speech detection
proficient to big data on spark.

4.2 Model Design and Classification
To develop the model, Spark ML API (spark.ml) which provides ML pipelines (workflow) for
creating, tuning, and evaluating of machine learning model was utilized. In Spark ML, a pipeline
is defined as a sequence of stages, and each stage is either a Transformer or an Estimator. These
stages are run in order, and the input DataFrame with spars vectors were transformed as it passes
through each stage.
Figure 3: Spark ML pipeline for training (Adopted from Apache spark)
Annotated data were given to the pipeline to get features as feature vectors. The study split the
dataset into two datasets, 80% (4882, comments) as training dataset and 20% (1238, comments)
as testing dataset using the spark DataFrame random split function with the seed of 100. The
training dataset was used to train model, and test dataset was used to evaluate the model
performance.
Figure 4: Spark ML pipeline for testing (Adopted from Apache spark)
4.3 Model Evaluation
For accuracy of the model, authors used cross validation using Spark evaluation tool namely
Binary class Classification Evaluator within the spark ML. To evaluate the performance of the
model classification in terms of quality or predictive effectiveness, different metrics appropriate
for the work accuracy, ROC score and Area under curve F-measure (F1-score) were used as
shown in table 3.

Table 3: Classification Performance result
Classifier Algorithm Feature Model
Evaluation Metrics Result
Accuracy ROC score Area under PR
Naive Bayes TF-IDF 0.73021 0.8053 0.7993
Word2Vec 0.7983 0.8305 0.8534
Random Forest TF-IDF 0.6355 0.6844 0.6966
Word2Vec 0.6534 0. 7097 0.7307
4.4 Results and Analysis
Authors evaluated classification model by using the 10-fold cross-validation method, achieving
an average result as presented in table 3. It was evident that the Naïve Bayes classifier with
word2Vec feature model outperform to classify hate and Nohate speech 0.7983, 0.8305 and
0.8534 accuracy, ROC score and area under Precision and Recall respectively with Facebook
social network for Amharic language posts and comments. The Naïve Bayes also achieve better
result for TF-IDF feature model with 0.73021, .08053 and 0.7993 for accuracy, ROC score and
area under precision and recall respectively. The Random Forest with word2vec feature is better
than TF-IDF with the result 0.6534, 0.7097 and 0.7307 accuracy, ROC score and area under
precision and recall respectively. This is followed by TF-IDF with 0.6355, 0.6844 and 0.6996
respectively. Even though may not be appropriate to compare the result with different
experimental setups authors got the state of the art result found in other languages with unique
feature of scalability for big data.
The following two charts shows sample of the hate speech classification performance using ROC
score area.
Figure 5: ROC for Naïve Bayes with TF-IDF

Figure 6: ROC for Random Forest with Word2Vec
5. CONCLUSIONS AND FUTURE WORK
The study developed a model for Amharic text hate speech detection that analyzes posts and
comments to identify hate speech using spark machine learning techniques. To conduct the
experiments, thousands of Amharic post and comments on suspected social network pages of
organizations and individual person’s public pages are crawled as dataset. First prepressed
according to the requirement of the language and human annotators selected to label the comment
in to hate or not hate. Here after, features are pipelined to word2vec neural network tool and TF-
IDF in apache spark environment so that feature vectors are obtained.
The classification algorithms were implemented in Apache Spark local cluster using the Apache
Spark’s Machine Learning library. The model developed using Naïve Bayes and Random forest
utilizing a dataset of 6,120 Amharic posts and comments out of this 4,882 to train the model and
1,238 for testing after passing different steps as stated in the experiment section. The model was
tested to classify whether the post and comments are hate or not and able to detect and classify in
an accuracy of 79.83 % and 65.34% for Naïve Bayes with word2vec feature vector and Random
Forest with TF-IDF feature modeling approach respectively. The workshow that word2vec
feature model is better in maintaining the semantics of the posts and comments as proved in other
works. The result are promising for such work in social network big data which can be extended
to compute large volumes data since the work used the distributed platform of apache spark.
Even if the results are promising for hate detection, our research is far from perfect. A lot of work
ahead of us to work on technical improvements that can be made for the language interms of: (1)
expand the dataset that would reduce the risk of overfitting and improve the statistical
significance of the results (2)analyzing the different aspect of the category of hate, either hate
with politics, ethnicity, religion and socio-economy (3)utilize the information provided by
Facebook so that, classification can be improved by expanding the feature space with profile
information, list of followers and geolocation etc. (4) crawl other sources to improve the feature
space for such under resourced language for computational purpose by adding synonyms from
other sources such as Twitter, forums and other homepages.

Finally, the proposed methods could be applied in different domains where the posts about the
anticipation to get service and buy product by the review of the service after serving or buying it
for this particular language showing the sentiment of costumers as positive or negative can be
explored.
REFERENCES
[1] Bjorn Gamback and Utpal Kumar Sikdar, Named Entity Recognition for Amharic Using Deep
Learning. IST-Africa 2017 Conference Proceedings Paul Cunningham and Miriam Cunningham
(Eds) IIMC International Information Management Corporation, ISBN: 978-1-905824-57-1, 2017
[2] META-NET White Paper Series, Retrieved from Multilingual Europe Technology Alliance:
http://www.meta-net.eu/whitepapers/overview , (2018, January Tuesday)
[3] Grover Hudson, Linguistic analysis of the 1994 Ethiopian census, Northeast African Studies, 6(3):89–
107, 1999
[4] Raphael Cohen-Almagor. Internet History, International Journal of Techno ethics, Vol. 2, No. 2, pp.
45-64, 2011
[5] Waseem & Hovy, Hateful Symbols or Hateful People? Predictive Features for Hate Speech Detection
on Twitter. Proceedings of NAACL-HLT, 2016, pages 88–93
[6] White Paper series, Retrieved from Amnesty International, Social media and Internet:
https://www.amnesty.org/en/latest/news/2016/12/ethiopia-government-blocking-of-websites-during-
protests-widespread-systematic-and-illegal/, 2016, 2018.
[7] Saleem, Haji Mohammad, Kelly P. Dillon, Susan Benesch, and Derek Ruths, A web of hate: Tackling
hateful speech in online social spaces. ArXiv preprint arXiv: 1709.10159, 2017.
[8] Fortuna, Paula Cristina Teixeira, Automatic detection of hate speech in text: an overview of the topic
and dataset annotation with hierarchical classes, 2017.
[9] Kwok, Irene, and Yuzhou Wang, Locate the Hate: Detecting Tweets against Blacks, In AAAI. 2013
[10] Del Vigna12, Fabio, Andrea Cimino23, Felice Dell’Orletta, Marinella Petrocchi, and Maurizio
Tesconi, Hate me, hate me not: Hate speech detection on Facebook, 2017.
[11] Silva, Leandro Araújo, Mainack Mondal, Denzil Correa, Fabrício Benevenuto, and Ingmar Weber,
Analyzing the Targets of Hate in Online Social Media, In ICWSM, pp. 687-690, 2016.
[12] Waseem, Zeerak, and Dirk Hovy, Hateful symbols or hateful people? Predictive features for hate
speech detection on twitter, In Proceedings of the NAACL student research workshop, pp. 88-93,
2016.
[13] 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, pages 145-153. International World Wide Web Conferences Steering Committee, 2016.
[14] Davidson, Thomas, Dana Warmsley, Michael Macy, and Ingmar Weber, Automated hate speech
detection and the problem of offensive language, arXiv preprint arXiv: 1703.04009, 2017.

[15] Yashar Mehdad and Joel Tetreault, Do characters abuse more than words? In Proceedings of the
SIGdial 2016 Conference: The 17th Annual Meeting of the Special Interest Group on Discourse and
Dialogue, pages 299–303, 2016.
[16] Keith Cortis and Siegfried Handschuh, Analysis of cyberbullying tweets in trending world events. In
Proceedings of the 15th International Conference on Knowledge Technologies and Data-driven
Business, page 7. ACM, 2015.
[17] Agarwal, Swati, and Ashish Sureka, Characterizing Linguistic Attributes for Automatic Classification
of Intent Based Racist/Radicalized Posts on Tumblr Micro-Blogging Website, arXiv preprint arXiv:
1701.04931, 2017.
[18] Mikolov, Tomas, Kai Chen, Greg Corrado, and Jeffrey Dean, Efficient estimation of word
representations in vector space, arXiv preprint arXiv: 1301.3781, 2013.
[19] Mikolov, Tomas, Ilya Sutskever, Kai Chen, Greg S. Corrado, and Jeff Dean, Distributed
representations of words and phrases and their compositionality, In Advances in neural information
processing systems, pp. 3111-3119, 2013.
[20] BjornRoss, MichaelRist, Guillermo Carbonell, Benjamin Cabrera, Nils Kurowsky, and Michael
Wojatzki, Measuring the reliability of hate speech annotations: The case of the European refugee
crisis, 2017.
[21] Stephan Tulkens, Lisa Hilte, Elise Lodewyckx, Ben Verhoeven, and Walter Daelemans, A
dictionary-based approach to racism detection in Dutch social media, 2016.
[22] 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, pages 86–95, 2017.
[23] Liu, Shuhua, and Thomas Forss, Combining n-gram based similarity analysis with sentiment analysis
in web content classification, In Proceedings of the International Joint Conference on Knowledge
Discovery, Knowledge Engineering and Knowledge Management-Volume 1, pp. 530-537.
SCITEPRESS-Science and Technology Publications, Lda, 2014.
[24] Liu, Shuhua, and Thomas Forss, New classification models for detecting Hate and Violence web
content, In Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K),
2015 7th International Joint Conference on, vol. 1, pp. 487-495. IEEE, 2015.
[25] Maloba, Wilson Jeffrey, Use of regular expressions for multi-lingual detection of hate speech in
Kenya, PhD diss., iLabAfrica, 2014.
[26] Njagi Dennis Gitari, Zhang Zuping, Hanyurwim fura Damien, and Jun Long. A lexicon-based
approach for hate speech detection, International Journal of Multimedia and Ubiquitous Engineering,
10(4):215–230, 2015.
[27] Nemanja Djuric, JingZhou, Robin Morris, Mihajlo Grbovic, Vladan Radosavljevic, and Narayan
Bhamidipati, Hate speech detection with comment embed dings. In Proceedings of the 24th
International Conference on World Wide Web, pages 29–ACM2, 2015.
[28] Pinkesh Badjatiya, Shashank Gupta, Manish Gupta, and Vasudeva Varma, Deep learning for hate
speech detection in tweets. In Proceedings of the 26th International Conference on World Wide Web
Companion, pages 759–760. International World Wide Web Conferences Steering Committee, 2017.

[27] Maral Dadvar, Franciska de Jong, Roeland Ordelman, and Dolf Trieschnigg, Improved cyberbullying
detection using gender information. In Proceedings of the Twelfth Dutch-Belgian Information
Retrieval Workshop, pages 23–25, University of Ghent, 2012.
[28] Shuhan Yuan, Xintao Wu, and Yang Xiang, A two phase deep learning model for identifying
discrimination from tweets, In International Conference on Extending Database Technology, pages
696–697, 2016.
[29] Kwok and Wang, Detecting Tweets against Blacks. Proceedings of the Twenty-Seventh AAAI
Conference on Artificial Intelligence, 2013
[30] William Warner and Julia Hirschberg, Detecting Hate Speech on the World Wide Web. Proceedings
of the 2012 Workshop on Language in Social Media (LSM 2012), pages 19–26, Association for
Computational Linguistics, Montreal, Canada, 2012
[31] Anna Schmidt and Michael Wiegand, A survey on hate speech detection using natural language
processing. Social NLP 2017, page 1, 2017.
[32] Anthony J. Viera, Understanding Inter observer Agreement: The Kappa Statistic, From the Robert
Wood Johnson Clinical Scholars Program, University of North Carolina, 2005
[33] Kilem L Gwet. Handbook of inter-rater reliability: The deﬁnitive guide to measuring the extent of
agreement among raters. Advanced Analytics, LLC, 2014.
[34] Joseph Lilleberg et al, Support Vector Machines and Word2vec for Text Classification with Semantic
Features. Proc. 20151IEE 14th Inl'l Coni. On Cognitive Informatics & Cognitive Computing IEdsJ,
2015.
[35] United Nations Educational, Scientific and Cultural Organization, Countering Online Hate Speech.
Published in 2015 by the United Nations Educational, Scientific and Cultural Organization 7, place de
Fontenot, 75352 Paris 07 SP, France
[36] Iginio Gagliardone, Alisha Patel and Matti Pohjonen, Mapping and Analyzing Hate Speech
Online:Opportunities and Challenges for Ethiopia, 2014

AppendixA
Experiment setup
Apache environment setup
Apache Spark-2.2.0 Ubuntu 16.4 virtual machine
Standalone clustering mode
1 master node
2 worker node
Feature selection Algorithms
Name of Algorithm Parameter set up
Word2Vec choice of training model
0: Skip gram model
dimension of vectors=3
Window size=5
Minimum count=0
TF-IDF Default
Machine learning Algorithms
Naïve Bayes Model type =multinomial
Smoothing =1
Random Forest NumTree=200
MaxDepth=3
Seed=2
10-fold cross-validation
Naïve Bayes Smoothing =1
Numfolds=10
Random Forest NumTree=[50,100,200]
MaxDepth=[3,4,5]
Numfolds=10

SOCIAL NETWORK HATE SPEECH DETECTION FOR AMHARIC LANGUAGE

More Related Content

What's hot (20)

Similar to SOCIAL NETWORK HATE SPEECH DETECTION FOR AMHARIC LANGUAGE (20)

More from cscpconf (20)

Recently uploaded (20)

SOCIAL NETWORK HATE SPEECH DETECTION FOR AMHARIC LANGUAGE