XXXVII Congresso da Sociedade Brasileira de Computação

Offensive Comments in the Brazilian Web: a dataset and

baseline results
Rogers Prates de Pelle, Viviane P. Moreira
1
Instituto de Informática – Universidade Federal do Rio Grande do Sul (UFRGS)
Caixa Postal 15.064 – 91.501-970 – Porto Alegre – RS – Brazil
{rppelle,viviane}@inf.ufrgs.br

Abstract. Brazilian Web users are among the most active in social networks
and very keen on interacting with others. Offensive comments, known as hate
speech, have been plaguing online media and originating a number of law-
suits against companies which publish Web content. Given the massive number
of user generated text published on a daily basis, manually filtering offensive
comments becomes infeasible. The identification of offensive comments can be
treated as a supervised classification task. In order to obtain a model to clas-
sify comments, an annotated dataset containing positive and negative examples
is necessary. The lack of such a dataset in Portuguese, limits the development
of detection approaches for this language. In this paper, we describe how we
created annotated datasets of offensive comments for Portuguese by collecting
news comments on the Brazilian Web. In addition, we provide classification re-
sults achieved by standard classification algorithms on these datasets which can
serve as baseline for future work on this topic.

1. Introduction
Hate speech is defined as “any communication that disparages a person or a group on
the basis of some characteristic such as race, color, ethnicity, gender, sexual orientation,
nationality, religion or other characteristic” [Nockleby, 2000]. Recently, companies like
Facebook, Twitter, and YouTube have been facing legal action for allowing users to post
texts that are considered offensive1 . Since there is a massive number of posts every day,
there is a need for automatically identifying and filtering such posts.
In Brazil, there have been prominent cases of racist texts posted in social networks
targeting black celebrities23 . Reports say that Brazilians spend the most time on social
media and estimate that 96% of all Brazilian Internet users have at least one social network
account [Webcertain, 2015]. Facebook alone has over 100 million Brazilian accounts4 .
The country ranks at third place in number of Facebook users and fifth in the number of
Twitter users5 .
Hate speech is not limited to social networks. In a preliminary analysis, we took a
sample of 145 news published on a single day (7-Jun-2016) by the biggest news site in the
1
http://www.bbc.com/news/technology-36301772
2
http://glo.bo/1Hl0HaB
3
http://glo.bo/1UiV3i0
4
http://www.internetworldstats.com/south.htm
5
http://www.forbes.com/sites/victorlipman/2014/05/24/
top-twitter-trends-what-countries-are-most-active-whos-most-popular

510
 6º BraSNAM - Brazilian Workshop on Social Network
Analysis and Mining

country6 and identified that 90% had at least one hateful comment. In most cases, users
start discussing the news and end up engaging in arguments using abusive language. Even
though companies have mechanisms for preventing the publication of offensive texts, they
still are not able to prevent cases such as the aforementioned ones.
In Brazil, the company Agência Nova/SB7 carried out a study on intolerance in so-
cial networks. During three months, they monitored Facebook, Twitter, Instagram, and a
number of blogs and websites. They recorded every time a subject such as racism and ho-
mophobia was mentioned, collecting a total of 542,781 mentions. Their findings, reported
in [NoavaS/B, 2016], state that 84% of the comments on these subjects are negative.
Identifying offensive comments is not easy because authors tend to disguise of-
fensive words by inserting asteriscs, spaces, or replacing characters by others with similar
sounds. Thus, simply checking for the presence of terms that are in a precomputed list of
offenses (i.e., a blacklist) would miss many such comments. A better solution would be
modeling this task as a text classification problem and generate models that would learn
automatically how to identify offensive comments. This approach requires an annotated
dataset with positive and negative examples (i.e., offensive and non-offensive comments).
A number of recent works addressed the detection of hate speech on the Web fo-
cusing on different aspects, such as identifying racist tweets [Kwok and Wang, 2013, Silva
et al., 2016] or blogs [Chau and Xu, 2007, Warner and Hirschberg, 2012], filtering pages
with hate and violence [Liu and Forss, 2015], detecting hate groups [Ting et al., 2013],
identifying flames [Razavi et al., 2010], and offensive news comments [Sood et al., 2012,
Djuric et al., 2015]. To the best of our knowledge, none of such works have addressed the
Brazilian Web or Portuguese texts.
The contributions of this work are: (i) an annotated dataset containing offensive
(and non-offensive) comments collected from the Brazilian Web. We call it O FF C OM B R
and make it available to the research community; (ii) baseline results of standard classifi-
cation algorithms applied to the dataset, which may serve as reference for future works on
this topic; and (iii) as a byproduct of the annotation process, we made available a system
to help judges annotate sentences. We believe our contributions represent an initial step
to enable the development of the identification of hate speech in Portuguese.
This paper is organized as follows. Section 2 presents the related literature on
hate speech identification. Section 3 describes the process used to create O FF C OM B R
and presents statistics on the dataset. Section 4 reports on experimental results on the
datasets. Section 5 discusses our main findings; and Section 6 concludes the paper.

2. Related Work
The high volumes of hate speech on the Web has motivated the research community to
come up with approaches to identify such offenses. Recent work have addressed different
platforms such as web pages, social networks, blogs, and tweets.
Supervised classifiers have been used in a number of approaches for hate speech
detection in different platforms: Xiang et al. [2012] applied them to a dataset of tweets
looking for profane language; Kwok and Wang [2013] also worked with tweets but to
6
www.g1.globo.com
7
http://www.novasb.com.br

511
XXXVII Congresso da Sociedade Brasileira de Computação

detect racist posts; Razavi et al. [2010] used them for flame detection on newsgroup mes-
sages; Warner and Hirschberg [2012] tackled anti-semitic comments from Yahoo! groups;
and, while Sood et al. [2012] focused on detecting profanity, Djuric et al. [2015] addressed
the broad category of hate speech in news comments.
Works that created annotated datasets. Supervised learning depends on the existence
of annotated datasets containing instances with and without hate speech. A few studies
report having created annotated datasets using human judges. Sood et al. [2012] used
crowdsourcing to annotate 6,500 news comments. A total of 221 annotators identified
9,4% as containing profanity with at least 66% consensus. Warner and Hirschberg [2012]
had three judges label 1,000 paragraphs extracted from web pages that might contain anti-
semitic contents. Djuric et al. [2015] reports working with over 800K comments extracted
from the Yahoo! Finance website, but the annotation process is not clear. Nobata et al.
[2016] assembled a corpus with over 2 million comments from Yahoo! Finance and News.
The annotations were carried out by trained employees. Chen et al. [2012] asked judges
to assess whether 249 YouTube users were being offensive in their posts. Unfortunately,
we found that none of the aforementioned datasets are readily available.
Available Datasets. Despite the research on hate speech detection dating back to 20
years, only in 2016 did the first datasets become publicly available. Ross et al. [2016]8 has
470 tweets in German classified as to whether they contain hate speech against refugees.
Wulczyn et al. [2016]9 contains 115,737 Wikipedia discussion comments in English. The
annotations indicate whether the comment has a personal attack. Waseem [2016]10 has
6,909 tweets in English annotated for hate speech through a crowdsourcing effort. We also
found a dataset with about 4K English tweets classified as offensive and non-offensive on
the Kaggle website11 .
None of the datasets available are in Portuguese. Thus, the dataset we describe in
the next Section is, to the best of our knowledge, the first initiative in this direction.

3. Dataset Creation
In this section, we detail how the dataset was collected, the annotation process, and the
statistics of our datasets.
3.1. Data collection
The source of the data was the news site g1.globo.com. This is the most accessed
news site in Brazil12 and, as a result, it has many comments. Although the comments on
this site go through moderation, we found a considerable number of offensive contents.
About 90% of the news we analyzed had at least one offensive comment. After a prelim-
inary analysis, we noticed that the news categories with the most offensive comments are
politics and sports. Thus, our data collection was limited to those sections.
We implemented a webscraper which sends requests to the web site on the sections
of interest. The HTML pages of the news are then downloaded and parsed to extract the
8
https://github.com/UCSM-DUE/IWG_hatespeech_public
9
https://figshare.com/articles/Wikipedia_Detox_Data/4054689
10
https://github.com/zeerakw/hatespeech
11
https://www.kaggle.com/c/detecting-insults-in-social-commentary/
data
12
http://www.alexa.com/topsites/countries/BR

512
 6º BraSNAM - Brazilian Workshop on Social Network
Analysis and Mining

text of the news and the link where all comments for a given news can be obtained in
the JSON format. All comments for a given news were extracted. Comments that were
composed only of emojis or other non-alphabetical characters were discarded. To help
preserve the anonymity of the authors, whenever a comment mentioned the full name of
the author of another comment, we kept only the first name. At the end of this process,
we obtained 10,336 comments posted for 115 news.

3.2. Annotation Process

Since our bottleneck for creating the dataset is the availability o human judges, we could
not label all 10K instances. Thus, a sample of 1,250 comments was randomly selected.
Following the standard procedure adopted for dataset annotation, each comment was an-
notated by three judges which were asked to whether it was offensive. In case of an affir-
mative answer, the annotator was also asked to categorize the offence as racism, sexism,
homophobia, xenophobia, religious intolerance, or cursing.
We developed a tool to help the judges during the annotation process. The Web
interface (depicted in Figure 1) showed the comment, the categories, and a link to the
news for which the comment was written. A help screen with definitions of the cate-
gories and a definition of offensive text were also provided. We believe this tool could be
useful to other annotation tasks, so we made it available at http://inf.ufrgs.br/
˜rppelle/hatedetector/.
Two datasets were generated from the annotations. The first, called O FF C OM B R-
2, has all 1,250 instances and the class assigned to each comment was the one picked
by at least two of the judges. The second is a more strict dataset, called O FF C OM B R-3.
This dataset was composed solely of the comments for which all three judges agreed as
to whether or not the comment was offensive. The datasets can be obtained at http:
//inf.ufrgs.br/˜rppelle/hatedetector/.

Figure 1. Interface of the annotation tool

513
XXXVII Congresso da Sociedade Brasileira de Computação

3.3. Statistics for the datasets

To measure the level of agreement among the judges, we calculated the Fleiss
Kappa Fleiss [1971] measure which quantifies degree of agreement over that which would
be expected by chance. For O FF C OM B R-2, the value was 0.71, which is considered sub-
stantial. This value is within the range of agreement found in other works that also carried
out annotations (0.63 for Warner and Hirschberg [2012], 0.73 for Chen et al. [2012], and
0.84 for Nobata et al. [2016]). Since O FF C OM B R-3 only contains instances for which the
class has been agreed by all three judges, it made no sense to calculate Kappa.
In O FF C OM B R-2, 419 (out of 1,250) comments were considered offensive by at
least two judges, representing 32,5% of the total (we noticed that no comment was found
offensive by only one judge)13 . For O FF C OM B R-3, there are 202 offensive comments (out
of 1,033), amounting to 19,5% of the cases. As with other datasets for hate speech detec-
tion, both versions of O FF C OM B R are unbalanced with negative examples outnumbering
positive ones.
Regarding the categories (racism, sexism, homophobia, xenophobia, religious in-
tolerance, or cursing), the results for each dataset are shown in Table 1. The most common
category was by far cursing. The other categories had few comments. Religious intoler-
ance was found in only one comment and it was not unanimous.

Table 1. Prevalence of each Category in the Annotations

Religious
# Judges Xenophobia Homophobia Sexism Racism Cursing
Intolerance
1 13 (1,0%) 35 (2.8%) 14 (1,1%) 19 (1.5%) 375 (30.0%) 1 (0.1%)
2 12 (1.0%) 14 (1,1%) 8 (0.6%) 18 (1.4%) 286 (22.9%) 1 (0.1%)
3 5 (0.5%) 9 (0.9%) 4 (0.4%) 1 (0.1%) 175 (16,9%) 0 (0.0%)

4. Implementing Classifiers to Identify Offensive Comments

In this section, we address the identification of offensive comments as a text classification
problem. The task was then to take each comment perform a binary classification as to
whether it is offensive (i.e.,, the classes were yes and no). Our goal here is to provide
baseline results of standard classification algorithms and data preprocessing tasks on our
datasets.

4.1. Experimental Setup

The instances in both datasets were submitted to a number of standard preprocessing
tasks. These are identified below:
• Case folding. Two options were adopted: (i) converting the text of the comments
to lowercase (lower) and (ii) leaving the comments in the case it was typed by
the authors (original).
13
Note that here we report on the percentage of comments that were found offensive, while in the Intro-
duction we did a preliminary analysis of the percentage news articles with offensive comment. We found
that, while 90% of the news articles had at least one offensive comment, out of all comments, 32,5% were
considered offensive.

514
 6º BraSNAM - Brazilian Workshop on Social Network
Analysis and Mining

Table 2. Statistics of the two datasets

O FF C OM B R-2 O FF C OM B R-3
File #features file #features
original 1G 4,980 original 1G 4,348
original 1G FS 241 original 1G FS 119
original 1G+2G 17,374 original 1G+2G 15,085
original 1G+2G FS 396 original 1G+2G FS 164
original 1G+2G+3G 30,711 original 1G+2G+3G 26,600
original 1G+2G+3G FS 448 original 1G+2G+3G FS 182
lower 1G 4,123 lower 1G 3,647
lower 1G FS 250 lower 1G FS 122
lower 1G+2G 15,899 lower 1G+2G 12,385
lower 1G+2G FS 426 lower 1G+2G FS 103
lower 1G+2G+3G 29,126 lower 1G+2G+3G 25,303
lower 1G+2G+3G FS 489 lower 1G+2G+3G FS 196

• Tokenization. The text of the comments is tokenized and the tokens are used as
features by the classification algorithms. This is known as bag-of-words approach.
Three options of extracting n-gram (sequences of n tokens) features from the texts
were tested: (i) using unigrams (1G), (ii) using unigrams and bigrams (1G+2G),
(iii) using unigrams, bigrams and trigrams as features (1G+2G+3G). The idea is
that by using longer n-grams, we could capture better the structure of the dis-
course.
• Feature selection. we compared the use of all n-grams as features against using
only the features selected by Information Gain. Information Gain measures how
correlated each feature is with respect to the class we wish to predict. We wanted
to keep only features that had a positive correlation with the class, thus the thresh-
old of zero was applied to the InfoGainAttributeEval method in Weka.
Combining all possibilities for case folding, tokenization, and feature selection
across both datasets yielded 24 files to be processed by the classification algorithms.
Statistics for these files are shown in Table 2.
We tested two algorithms that are widely used for text classification: Naive Bayes
and SVM (called SMO in Weka’s implementation). A 10-fold-cross-validation approach
was followed and the results we report on the next section are the averages of the ten
executions of each algorithm over the 24 data files.

4.2. Results
Our analysis of the results is based on the macro weighted average F-measure (i.e.,, the
weight is given by the class size). For each class, the F-measure is the harmonic mean
between recall and precision. In order to test whether the different performances were
statistically significant, we used paired T-tests with the F-measures of the executions with
the standard threshold of statistical significance of α = 0.05.
Classification results are shown in Figure 2. The columns reflect the average F-
measure across all executions for a given parameter (i.e., choices for preprocessing and

515
XXXVII Congresso da Sociedade Brasileira de Computação

0.85 0.82 0.82

0.81 0.80 0.81 0.80
0.79 0.78 0.79
0.80 0.77
F-measure

0.75 0.76
0.74 0.74 0.74
0.75 0.73 0.72 0.71
0.70

0.65
Yes

No

Yes

No
1G

1G+2G

1G+2G+3G

SMO

1G

1G+2G

1G+2G+3G

SMO
lower

original

NB

lower

default

NB
Case FS Tokenizer Classifier Case FS Tokenizer Classifier
OFFComBr-2 OFFComBr-3

Figure 2. Classification results for different data preparations

classification algorithm). For example, the first column shows the average of the weighted
F-measure in all executions in which the text of the comments was converted to lowercase
considering both classification algorithms. Absolute values for F-measure ranged from
.69 to .85. The configuration with the best result, in both O FF C OM B R-2 and O FF C OM B R-
3, was achieved by SMO over the files in which the comments converted to lowercase and
when feature selection was applied. The only difference was regarding the tokenization
options – 1G+2G was the best in O FF C OM B R-2 and 1G was the best in O FF C OM B R-3,
however, the differences were very small.
Regarding the results for each of the preprocessing tasks we found that:
• Case folding. Converting the text of the comments to lowercase brought signifi-
cant improvements, with the F-measure for lower being statistically superior to
the results for original. This suggests that the loss of information incurred by
case folding is compensated by a gain of generality in the classification model.
• Tokenization. We found no statistical difference among the three tokenization op-
tions. Considering bigrams increases features by a factor of three, and considering
trigrams increases the number further by a factor of two. Given the similar results
that the three alternatives yield and the additional cost in processing this many
more features, using unigrams only is preferable.
• Feature selection. Feature selection had a positive impact on the classifiers. Not
only did the F-measure get significantly better but also processing time was dras-
tically reduced by the use of much fewer features. The number of features was
reduced to a maximum of 489 in O FF C OM B R-2 and 196 in O FF C OM B R-3. This
represents a dramatic reduction as the selected features account for 0.8% to 4.8%
of the total features.
With respect to the classifiers, SMO performed significantly better than Naive
Bayes. The average F-measure for SMO was 0.80, while for Naive Bayes the average
was 0.75. Besides the difference in the scores, an important difference between the two
algorithms appears when we look at the misclassified instances. While SMO has more
false negatives (i.e., 188 offensive comments that were classified as non-offensive) and
only a few false positives (i.e., 15 non-offensive comments that were classified as offen-
sive), Naive Bayes has fewer false negatives (154) and many more false positives (184).
Although these results cannot be directly compared to other works that used su-
pervised classification algorithms for hate speech detection (since the datasets were dif-

516
 6º BraSNAM - Brazilian Workshop on Social Network
Analysis and Mining

ferent), we found that we have reached similar scores. Warner and Hirschberg [2012] and
Sood et al. [2012] report achieving 0.63, while Chen et al. [2012] and Nobata et al. [2016]
report scores of 0.8 and 0.83, respectively.
As expected, the classifiers performed better with O FF C OM B R-3 than with O FF -
C OM B R-2, since the difference between the yes and no classes are clearer in the former.

5. Discussion
In this Section we describe the main findings of this study and their implications.
Prevalence of cursing. Insults including vulgar language were the most frequent cate-
gory of offensive comments. They were present in almost 70% of the comments found
offensive. The targets of the insults were either people mentioned in the news (politicians
or soccer players) or authors of other comments.
The importance of context. In many cases, we found that the same words present in
offensive comments could also be used in non-offensive text. The distinction is the context
in which they are used. For example, in all the words in the comments “a minha mãe há
anos esta no céu mas a sua tá na zona” and “voce nasceu assim ou bateu a cabeca
quando era crianca” are also very common in non-offensive comments. This means that
one needs to know the context in which the words appear to be able to accurately classify
them. In our experiments, we tested the use brigrams and trigrams as features hoping
that they would provide more context and thus achieve better classification performance.
Our results, however, did not improve with the use of longer n-grams. This suggests
that other forms of providing context, such as considering sentence structure, should be
investigated.
Language is constantly changing. The language on the Web is full of jargon, mis-
spellings, and abbreviations. With the increase on the number of users, the evolution of
the language becomes faster. New words are created and old words gain new meanings.
This makes the use of static black-lists of offensive words inadequate. Also, if machine
learning classifiers are to be used, they need to be capable of adapting and continue to
learn new patterns.
Size of the comments. We observed a tendency that longer comments tended to contain
fewer offenses. This suggests that longer comments require more careful elaboration and
present fewer insults compared to shorter comments.
Freedom of speech. There is a fine balance between filtering offensive comments and
interfering with people’s freedom of speech. This calls for careful consideration in the
implementation of automatic methods which should aim to minimize false positives.
Limitations. Given that the availability of human judges to annotate the instances is
the bottleneck in the creation of datasets, O FF C OM B R is limited to 1,250 instances. We
feel that many more comments are needed to allow for a better coverage of the offensive
language identified on the Web. A larger dataset would also help classifiers distinguish
between the categories. In addition, the low prevalence of comments with racism, sexism,
homophobia, xenophobia, and religious intolerance makes O FF C OM B R not suitable for
research that focuses on these categories. Furthermore, since we did not evaluate every
comment for a given news article, we are unable to analyze whether articles with more
comments tend to have more offensive comments.

517
XXXVII Congresso da Sociedade Brasileira de Computação

6. Conclusion
Methods for hate speech detection that rely on supervised machine learning require anno-
tated datasets. Although a number of studies have been conducted in this topic, there are
only a handful of datasets available, most of them in English.
This paper addresses the lack of hate speech datasets in Portuguese by describing
the creation of annotated datasets of offensive posts collected from news comments. The
datasets, called O FF C OM B R-2 and O FF C OM B R-3, are freely available to the research
community.
We have also run standard classification algorithms on the datasets to provide
baseline results. The F-measure scores we obtained are within the range found in other
work that addressed hate speech identification in English.
Out future work will include enlarging our datasets through a crowdsourcing ef-
fort using a larger number of volunteer annotators. In addition, we plan on adding an
(anonymized) userId to each comment in the dataset. This will enable finding out the
prevalence of offensive comments by user. This could be used as additional evidence by
automatic methods for filtering offensive text.

Acknowledgements: This work has been partially supported by CNPq. R. P. de Pelle

receives an MSc grant from CNPq. We thank the volunteers Cássio Alan Garcia, Diego
de Vargas Feijó, Geisiane Martini, Lucas Pessutto, and Paula Burguêz for their effort in
annotating our datasets.

References
Michael Chau and Jennifer Xu. Mining communities and their relationships in blogs: A
study of online hate groups. International Journal of Human-Computer Studies, 65(1):
57–70, 2007.
Ying Chen, Yilu Zhou, Sencun Zhu, and Heng Xu. Detecting offensive language in
social media to protect adolescent online safety. In Privacy, Security, Risk and Trust
(PASSAT), 2012 International Conference on and 2012 International Confernece on
Social Computing (SocialCom), pages 71–80, 2012.
Nemanja Djuric, Jing Zhou, Robin Morris, Mihajlo Grbovic, Vladan Radosavljevic, and
Narayan Bhamidipati. Hate speech detection with comment embeddings. In Proceed-
ings of the 24th International Conference on World Wide Web Companion, pages 29–
30, 2015.
J.L. Fleiss. Measuring nominal scale agreement among many raters. Psychological Bul-
letin, 76(5):378–382, 1971.
Irene Kwok and Yuzhou Wang. Locate the hate: Detecting tweets against blacks. In
Twenty-Seventh AAAI Conference on Artificial Intelligence, 2013.
Shuhua Liu and Thomas Forss. Text classification models for web content filtering and
online safety. In 2015 IEEE International Conference on Data Mining Workshop
(ICDMW), pages 961–968, 2015.
NoavaS/B. The Webcertain Global Search & Social Report, 2016.
URL http://www.comunicaquemuda.com.br/dossie/
quando-intolerancia-chega-as-redes/.

518
 6º BraSNAM - Brazilian Workshop on Social Network
Analysis and Mining

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, 2016.
John T. Nockleby. Hate speech. In Encyclopedia of the American Constitution (2nd
ed.,edited by Leonard W. Levy, Kenneth L. Karst et al., New York: Macmillan, 2000),
pages 1277–1279, 2000.
Amir H Razavi, Diana Inkpen, Sasha Uritsky, and Stan Matwin. Offensive language
detection using multi-level classification. In Advances in Artificial Intelligence, pages
16–27. 2010.
Björn Ross, Michael Rist, Guillermo Carbonell, Benjamin Cabrera, Nils Kurowsky, and
Michael Wojatzki. Measuring the Reliability of Hate Speech Annotations: The Case
of the European Refugee Crisis. In Proceedings of NLP4CMC III: 3rd Workshop on
Natural Language Processing for Computer-Mediated Communication, volume 17 of
Bochumer Linguistische Arbeitsberichte, pages 6–9, 2016.
Leandro Araújo Silva, Mainack Mondal, Denzil Correa, Fabrı́cio Benevenuto, and Ingmar
Weber. Analyzing the targets of hate in online social media. In Proceedings of the Tenth
International Conference on Web and Social Media, pages 687–690, 2016.
Sara Owsley Sood, Judd Antin, and Elizabeth F Churchill. Using crowdsourcing to im-
prove profanity detection. In AAAI Spring Symposium: Wisdom of the Crowds, vol-
ume 12, pages 69–74, 2012.
I Ting, Shyue-Liang Wang, Hsing-Miao Chi, Jyun-Sing Wu, et al. Content matters: A
study of hate groups detection based on social networks analysis and web mining. In
Proceedings of the 2013 IEEE/ACM International Conference on Advances in Social
Networks Analysis and Mining, pages 1196–1201, 2013.
William Warner and Julia Hirschberg. Detecting hate speech on the world wide web.
In Proceedings of the Second Workshop on Language in Social Media, pages 19–26,
2012.
Zeerak Waseem. Are you a racist or am i seeing things? annotator influence on hate
speech detection on twitter. In Proceedings of the First Workshop on NLP and Compu-
tational Social Science, pages 138–142, November 2016.
Webcertain. The Webcertain Global Search & Social Report, 2015. URL
http://internationaldigitalhub.com/en/publications/
the-webcertain-global-search-and-social-report-2015.
Ellery Wulczyn, Nithum Thain, and Lucas Dixon. Ex machina: Personal attacks seen at
scale. CoRR, abs/1610.08914, 2016.
Guang Xiang, Bin Fan, Ling Wang, Jason Hong, and Carolyn Rose. Detecting offensive
tweets via topical feature discovery over a large scale twitter corpus. In Proceedings of
the 21st ACM International Conference on Information and Knowledge Management,
CIKM ’12, pages 1980–1984, 2012.

519