Automatic Detection of Cyberbullying on Social Networks

based on Bullying Features

Rui Zhao Anna Zhou Kezhi Mao

[email protected] [email protected] [email protected]
School of Electrical and Electronic Engineering, Nanyang Technological University
Singapore 639798

ABSTRACT these platforms are also places where users experience bully-
With the increasing use of social media, cyberbullying be- ing as victims, bullies or bystanders. One study conducted
haviour has received more and more attention. Cyberbul- by national anti-bullying charity Ditch the Label in 2013,
lying may cause many serious and negative impacts on a has shown that two out of three 13-22 years old who were
person’s life and even lead to teen suicide. To reduce and surveyed have been victims of cyberbullying.1 As reported
stop cyberbullying, one effective solution is to automatically in [18], approximately 43% of teens once reported being bul-
detect bullying content based on appropriate machine learn- lied through social media. Another study also shows that
ing and natural language processing techniques. However, cyberbullying victimization rate ranges from 10% to 40%
many existing approaches in the literature are just normal [9].
text classification models without considering bullying char- Different from physical bullying, cyberbullying is ”behind-
acteristics. In this paper, we propose a representation learn- the-scenes” and ”24/7”. Even worse, the bullying messages
ing framework specific to cyberbullying detection. Based on left on the Internet will not vanish over time but contin-
word embeddings, we expand a list of pre-defined insulting uously bother other users. Therefore, the consequences of
words and assign different weights to obtain bullying fea- cyberbullying are even more far-rearching and severe than
tures, which are then concatenated with Bag-of-Words and those of physical bullying.
latent semantic features to form the final representation be- To detect cyberbullying content underlying huge volumes
fore feeding them into a linear SVM classifier. Experimental of posts on social media, a good solution is to develop ma-
study on a twitter dataset is conducted, and our method is chine learning-based automatic cyberbullying detection sys-
compared with several baseline text representation learning tem to categorize the information and generate reports if
models and cyberbullying detection methods. The superior any cyberbullying is detected, so that all the sensitive in-
performance achieved by our method has been observed in formation would be modified or erased at the first time,
this study. preventing Internet users from overexposure to undesirable
information. Natural Language Processing and Text Mining
techniques are commonly used in this solution. The frame-
CCS Concepts work of machine learning-based automatic cyberbullying de-
•Information systems → Data mining; tection includes two parts: Representation Learning for In-
ternet Messages and Classification. Each Internet message
Keywords is firstly transformed into a fixed-length vector and the clas-
sifier can be trained in the training corpus over the learned
Cyberbullying Detection, Text Mining, Representation Learn- feature space. Finally, the trained classifier can detect the
ing, Bag-of-Words, Word Embeddings existence of cyberbullying content in each new Internet Mes-
sage. Similar to other text categorization tasks, the core and
1. INTRODUCTION vital step is the first step: numerical representation learning.
The Internet technology has made a great impact on the In this field, many previous approaches utilize Bag-of-Words
communication and relationship among people. In partic- (BoW) model to represent text. BoW model is a classical
ular, social media platforms now gain their popularity for model in which document is regarded as a multi-set of the
people of a wide range of ages. Social media provides users words contained in it and modeled as a vector whose weights
not only a good platform for communication and informa- indicate the occurrence of words in the document. One ma-
tion sharing, but also an easy access to fresh news. However, jor limitation of BoW is that each feature corresponding to a
term is assumed to be independent to each other and fail to
Permission to make digital or hard copies of all or part of this work for personal or
capture semantic information. Feature extraction is always
classroom use is granted without fee provided that copies are not made or distributed conducted over the BoW features [19]. Besides BoW model,
for profit or commercial advantage and that copies bear this notice and the full cita- word embeddings, as one exciting fruit of deep learning in
tion on the first page. Copyrights for components of this work owned by others than NLP community, are able to capture semantic information
ACM must be honored. Abstracting with credit is permitted. To copy otherwise, or re-
publish, to post on servers or to redistribute to lists, requires prior specific permission behind words. In word embeddings, similar words are close
and/or a fee. Request permissions from [email protected]. to each other, the cosine similarity between word embed-
ICDCN ’16, January 04-07, 2016, Singapore, Singapore
1
c 2016 ACM. ISBN 978-1-4503-4032-8/16/01. . . $15.00 http://www.express.co.uk/news/uk/433733/
DOI: http://dx.doi.org/10.1145/2833312.2849567 69-of-youth-face-cyberbullying
dings reflect the semantic similarity between two words.
In cyberbullying detection, some bullying messages always
contain insulting or curse words, which can be regarded as
Final Representation
discriminative features. This semantic information can be
utilized to provide a robust text representation. Some pre-
vious efforts have been made to use these bullying words. Di-
nakar et.al utilized Linear Discriminative Analysis to learn
label specific features [4], which can be regarded as artificial Bag-of-words Latent Semantic Bullying
bullying words. In addition, based on prior knowledge, Na- Features Features Features
har et.al collected a list of bullying-like words and proposed
a weighted TF-IDF scheme by scaling these corresponding
bullying-like features with a factor of two [14]. Different
from a simple utilization of bullying terms in these previous Figure 1: Illustration of final representation in our
approaches, our work provides an elaborated use of bullying proposed EBoW model.
terms based on word embeddings.
In this paper, we proposed a new representation learning
method for cyberbullying detection named embeddings en- and combine them with BoW features to train a classifier
hanced Bag-of-Words model (EBoW). In EBoW, we firstly [4]. The length of label-specific features is limited to be
define a list of insulting words based on expert knowledge less than the class numbers, which hinders the performance
and linguistic resources, i.e., insulting seeds. Then, based on boost. Nahar et.al magnified the weights corresponding to
word embeddings, we extend insulting seeds to define bul- bullying words by two times [14]. This work shares a simi-
lying features. Different weights are assigned to bullying lar motivation with the construction of bullying features in
features based on the cosine similarity between word em- our model that bullying features should be enhanced. How-
beddings. Considering insulting seeds are expanded through ever, they did not consider the words’ semantics and the
word embeddings, the construction of insulting seeds do not scaling operation was quite arbitrary. In addition, Nahar
require much manual labor, which will be illustrated in the et.al [15] also adopted topic models including Probabilis-
section 3.3. Finally, we concatenate the learned bullying tic Latent Semantic Analysis (PLSA) and Latent Dirichlet
features with Bag-of-Words features and Latent semantic Allocation (LDA) to learn topics and performed feature se-
features to form the vector representation. lection, which is conducted over topics that feature under
This paper is organized as follows. In section 2, we firstly bullying-like topics are preserved. However, the determina-
introduce some related work. Then, we present our proposed tion of bullying-like topics lacks a general theoretical basis.
Embedding-enhanced BoW (EBoW) model as a representa- 2.2 Word2vec Embeddings
tion learning method for cyberbullying detection in Section
3. In Section 4, experimental results on a real twitter corpus Recently, Google introduced word2vec, which is an effec-
are illustrated and analyzed. Finally, concluding remarks tive and efficient word-embeddings tool [12, 13]. Based on a
are provided in Section 5. two-layer neural network language model, word2vec learns
vector representations for each word. The tool actually in-
cludes two separated models: Continuous Bag of Words
2. RELATED WORK (CBoW) and Skip-gram. The training goals of these two
Since our work adopts word2vec embeddings to learn rep- methods are reverse. CBoW tries to predict a word given
resentation for cyberbullying detection, we briefly review the surrounding words, while Skip-gram tries to predict a
some previous works in cyberbullying detection and word2vec window of words given a single word. Due to its surpris-
embeddings. ingly efficient architecture and unsupervised training proto-
col, word2vec can be trained over a large-scale unannotated
2.1 Cyberbullying Detection corpus with limited computational resources. The meaning-
The booming of social network leads to the extensive spread ful linguistic relationships between words can be encoded
of cyberbullying, which is a quite severe problem for children into learned word2vec embeddings.
and teenagers. Traditional studies of cyberbullying stand
more on a macroscopic view. Conducted by social scientists 3. EMBEDDINGS-ENHANCED
and psychologists, those studies focus on the statistics of cy-
berbullying and how to prevent them in a psychological way BAG-OF-WORDS MODEL
[11, 6, 8, 5]. As big social network service providers all offer In this section, we present our proposed representation
open APIs for academic research, instead of doing statisti- learning method for cyberbullying detection. The final rep-
cal study on limited sampled data, researchers are able to resentation consists of three parts including Bag-of-Words
access to much larger corpus by using data crawling, which features, latent semantic features and bullying features based
further drives the development of the computational study on word embeddings, as shown in Figure 1. These three
of cyberbullying based on machine learning and natural lan- kinds of features are presented as follows, in particular, the
guage processing techniques. One introductory work has details of bullying features generation.
been presented in [17], in which several NLP models such as
BoW, Latent Semantic Analysis (LSA) and Latent Dirichlet 3.1 Bag-of-Words Features
Allocation (LDA) are applied to detect bullying signals in To extract Bag-of-Words features, a vocabulary including
social media. Their results have verified the possibility of unigram and bigram is constructed firstly and the terms
automatic cyberbullying detection. Dinakar et.al used Lin- whose document frequencies are less than 2 are all ignored.
ear Discriminative Analysis to learn label specific features Different term weighting schemes including tf-idf and binary
ones can be applied here [3, 16]. In this paper, we adopt the
tf-idf weighting scheme. The tf-idf weight corresponding to
the i-th word in j-th document is calculated as follows:

N
wi,j = T Fi,j × log( ) (1)
DFi
where T Fi,j is the term frequency of the i-th word in j-th
document, DFi is the number of documents containing i-th
word and N is the number of documents.

3.2 Latent Semantic Features

Here, latent semantic features refer to the features ex-
tracted by Latent Semantic Analysis (LSA) [10]. In prin-
ciple, LSA applies Singular Value Decomposition (SVD) on
the term-document matrix that each column is the above
Bag-of-Words features. Then, the derived latent space is
spanned by dominant eigenvectors corresponding to large
eigenvalues. Each new feature is a linear combination of all
Figure 2: Word cloud visualization of our pre-
original features.
defined insulting words, i.e., insulting seeds.
3.3 Bullying Features
Insulting words can be pre-defined and extracted based on
our prior knowledge and other public linguistic resources, Different from the one-hot representation in Bag-of-Words
which are named insulting seeds in this paper. Then, we model, word embeddings use real-valued and low-dimensional
extend these insulting words automatically based on word vectors to represent semantics of words. In the vector space
embeddings. The insulting words and their corresponding behind well-trained word embeddings, similar words are placed
extended terms serve as bullying features. close to each other. For example, the words China and Sin-
gapore are close to each other, and the words beef and pork
3.3.1 Insulting Seeds are close to each other. The cosine similarity between word
Since some cyberbullying messages usually contain curse embeddings can represent the semantic similarity between
or insulting words, these words are good indications of the the two corresponding words. Since we focus on twitter
existence of bullying. Therefore, we select a list of insult- corpus here, we utilize a well-trained word2vec model on
ing words based on our prior knowledge and some external a large-scale twitter corpus including 400 million tweets [7].
linguistic resources2 . This list contains 350 words indicat- A visualization of some word embeddings based on PCA is
ing curse or negative emotions, such as nigga, bitch, fuck, shown in Figure 3. It is observed that curse words form dis-
slut, whore, twat etc. Then, we compare the word list with tinct clusters, which are also far away from normal words.
the unigram and bigram features of the corpus used in ex- Even insulting words are located at different regions due to
periments to obtain the intersection, which is regarded as different word usages and insulting expression.
insulting seeds in this paper. For example, the word cloud We extend the pre-defined insulting seeds based on word
visualization of insulting seeds of the twitter dataset used in embeddings. For each insulting seed, we select the top-h
experimental study is shown in Figure 2. As discussed in most-similar words in the vocabulary as extended bullying
the previous work [14], these features are discriminative for features. Based on word embeddings, the cosine similarity
cyberbullying detection. of the two embeddings can measure the similarity between
These insulting words are used to construct a vocabulary, these two words. It should be noted that the vocabulary
and each Internet message is mapped to a vector by count- of our corpus consists of unigrams and bigrams. For one
ing the occurrence times of each word appears. Different bigram wl wr , we simply use an additive model to derive the
term weighting schemes including tf-idf and binary ones are corresponding embedding as follows:
applied. However, such a direct use of these insulting words
could hardly achieve robust representation. The reasons are
two-fold. Firstly, due to the linguistic complexity and the v(wl wr ) = v(wl ) + v(wr ) (2)
synonymy, the obtained insulting words can not cover all To give a clear illustration, the top-10 similar terms to the
the possible bullying terms. Secondly, effect intensities of insulting seed : ”slut” and their cosine similarities scores are
different insulting words are different. The occurrence of shown in Table 1. These cosine similarity scores can be used
slut is more indicative of the existence of cyberbullying than to weight these extended features. And the weights of these
that of nerd, although these two words both express nega- insulting seeds are set to 1. These extended features and
tive affection. To overcome these two limitations, we employ insulting seeds compose the final bullying features. It should
word embeddings to extend these insulting words and weight be noted that one word may be similar to several insulting
them. seeds. To address the possible overlapping, we only keep the
highest cosine similarity of the word as its weight. Therefore,
3.3.2 Bullying Features Generation based on Word
the length of the final bullying features is less than or equal
Embeddings
to n ∗ h where n is the number of insulting seeds.
2
A collection of insulting words can be found in the website:
http://www.noswearing.com/dictionary 3.4 Feature Concatenation
Table 1: Top-10 similar terms to the insulting word: Non-Bullying Trace
slut. They are retrieved based on word embeddings. 1 Don't let your mind bully your body into believing it must carry the
Their corresponding cosine-similarity scores are also burden of its worries. #TeamFollowBack
shown. 2 Whether life's disabilities, left you outcast, bullied or teased,rejoice
Similar Words Cosine Similarity Scores and love yourself today, 'Cause baby, you were born this way
a slut 0.815 3 @USERNAME haha hopefully! Beliebers just bring a new
whore 0.738 meaning to cyber bullying
a whore 0.638
hypocrite 0.536 Bullying Trace
1 @RodFindlay been sent a few of them. Thought they could bully
bitch 0.508 me about. Put them right and they won't represent the client
puta 0.468 anymore!
nerd 0.455 2 He a bully on his block, in his heart he a clown
bully her 0.451 3 I was bullied #wheniwas13 but now I am the OFFICE bully!!
fat bully 0.440
bully nigga 0.435

1.2 Figure 4: Some examples from Twitter Datasets.

book
1.1 read
Three of them are non-bullying traces. And the
good
other three are bullying traces.
1.0 nice
shit
0.9 Table 2: Statistical properties of the Twitter
dataset.
0.8
Statistics Twitter
fuck bitch Sample No. 1762
0.7
Bullying Instances 684
0.6 Train/Test 5-fold
stupid slut nerd
whore
0.5
0.7 0.6 0.5 0.4 0.3 0.2 0.1 0.0

and post tweets, which are defined as the messages posted on

Twitter with a maximum length of 140-character. Twitter
Figure 3: Two dimensional visualization of our used has over 270 million active users, and 500 million tweets are
Twitter word embeddings via PCA. Displayed terms posted per day. Behind these amazing statistics, cyberbully-
include both bullying ones and normal ones. It ing has sadly become a common occurrence. According to a
shows that similar words are nearby vectors. previous study4 , there were about 100,000 abusive tweets a
week. Therefore, data crawled from Twitter is a good source
for cyberbullying research.
The Twitter dataset is composed of tweets crawled by
The three above kinds of features including Bag-of-Words
the public Twitter stream API. Each tweet contains at least
features, latent semantic features and bullying features are
one of the following keywords: bully, bullied, bullying. Re-
concatenated to form the final representation. When final
tweets are removed by excluding tweets containing the acronym
representation for each Internet message is obtained, a clas-
‘’RT‘’. Finally, 1762 tweets are sampled uniformly from
sifier is built to detect the existence of cyberbullying. In our
the whole tweets collections on August 6, 2011 and man-
proposed system, we feed learned features into a linear SVM
ually labeled. It should be pointed out here that labeling
pattern classifier.
is based on bullying traces. Bullying traces are defined as
response to the bullying experience, which include but far
4. EXPERIMENTS exceed the incidences of cyberbullying. Some examples of
In this section, we evaluate our proposed EBoW model bullying traces are shown in Figure 4. To preprocess these
using a real-world cyberbullying dataset. The details of our tweets, a twitter-specialized tokenizer5 is applied without
adopted dataset and experimental settings are introduced any stemming or stopword removal operations. In addition,
firstly. Experimental results and comparison with other some special characters including user mentions, URLS and
baseline methods are followed. so on are replaced by predefined characters, respectively.
The statistics of this dataset can be found in Table 2. Since
4.1 Descriptions of Datasets the dataset does not have explicit train/test split, 5-fold
The public bullying traces dataset is adopted here3 , which cross validation (CV) is applied, where four-folds are used
consists of tweets that are messages sent on Twitter [17]. for training and the remaining one is used for testing. The
Twitter is a very popular social media application, which is mean results will be reported.
able to let you access to the latest stories, ideas, opinions and
news(https://about.twitter.com/). Registered users can read 4.2 Experimental Setup
3 4
The dataset has been kindly provided at http://research. http://news.wisc.edu/20931
5
cs.wisc.edu/bullying/data.html http://www.ark.cs.cmu.edu/TweetNLP/
 The following methods will be compared: Table 3: Precision (%), Recall (%) and F1 Scores
(%) for compared methods. Bold face indicates best
* BoW Model: the raw BoW features are directly fed performance.
into the classifier. Measures BoW sBoW LSA LDA EBoW
Precision 75.6 75.7 75.9 74.0 76.8
* Semantic-enhanced BoW Model: This approach is re- Recall 77.8 78.3 78.2 76.5 79.4
ferred in [14]. Following the original setting, we scale F1 Score 76.6 76.9 77.0 74.9 78.0
the bullying features by a factor of 2. To give a fair
comparison, the bullying features here is insulting seeds
used in our proposed method. 80

* LSA: Latent Semantic Analysis [10].

79
* LDA: Latent Dirchilet Allocation [1]. Our implemen-
tation of LDA is based on Gensim 6 .

Scores (%)
78

* EBoW: Our proposed Embeddings-enhanced Bag-of-

Words Model. 77

For BoW, the most frequent 2000 terms including unigram

and bigram are used as features. For LSA and LDA, the 76

number of latent topics are both set to 100. To implement

F1 precision recall
LDA, we set hyperparameter α for document topic multino- 75
0 10 50 100 150 200 250
mial and hyperparameter η for word topic multinomial to 1 Number of Selected Similar Words for Each Insulting Seed
and 0.01, respectively.
For our proposed method, BoW and latent semantic fea-
Figure 5: Performance of our proposed EBoW
tures are both the same as those in two above compared
model for different h. Different colors indicate dif-
models. For bullying features, the number of insulting seeds
ferent measures.
is 20. The number of similar words compared to each in-
sulting seed h is set to 50. After feature expanding based
on word embeddings, the length of bullying features is 641 This may be because that both methods belong to dimen-
finally. Therefore, the dimension of features learned by our sionality reduction. The reduced dimension is a key param-
EBoW is the sum of the lengths of the three kinds of features eter to determine the quality of learned feature space. Here,
as: 2000 + 100 + 641 = 2741. we fix the dimension of latent space to 100. A deliberate
We then apply linear SVM [2] in the new feature space searching for this parameter may boost the performances of
generated by the above mentioned approaches. In linear LSA and LDA
SVM, the hyper-parameter: regularization term C, is searched
over a range as {2−2 , 2−1 , 20 , 21 , 22 , 23 , 24 , 25 } via a five-fold 4.4 Parameters Sensitivity
cross-validation conducted on training data. In our proposed EBoW model, we select the top-h similar
The cyberbullying detection task is thus converted into a words for each pre-defined insulting word. Parameter h is
binary classification problem. The adopted evaluation met- used as the threshold to expand bullying features. An in-
rics include precision, recall and F1-value. creased h will mean more terms are considered as bullying
features so that the dimension of bullying features increases.
4.3 Experimental Results In this section, we investigate the influence of the parame-
We reported the precision, recall and F1 scores of all com- ter on our model performance for cyberbullying detection.
pared methods in Table 3. Obviously, our proposed method Parameter h is chosen from a predefined set: [0, 50, 100,
outperforms other compared methods in all the three eval- 150, 200, 250]. When h = 0, bullying features only consist
uation metrics. of our pre-defined insulting words. According to the total 6
It is observed that semantic BoW model is able to give a settings of h, we feed our corresponding learned EBoW fea-
slight better performance than BoW. In sBoW, the weights tures into the classifier. The other experimental settings are
corresponding to bullying features are multiplied by 2 and kept unchanged. Then, precision, recall and F1 scores for
the bullying features are the same as the insulting seeds in these different settings are calculated and reported in Figure
our proposed EBoW. This indicates that bullying features 5.
can indeed improve cyberbullying detection. Different from It shows that a moderate h is able to achieve the best
a direct use of insulting seeds in sBoW, our approach EBoW performance. If h is too small, some discriminative terms
expands the pre-defined insulting words through word em- that are not covered by the pre-defined insulting words are
beddings and assign different weights to these expanded bul- filtered out in the learned bullying features. If h is too
lying features according to cosine similarity between word large, some irrelevant terms are considered as bullying fea-
embeddings. As a result, our approach is able to gain a sig- tures. Both scenarios will lead to an ineffective feature space
nificant performance improvement compared to sBoW over for cyberbullying detection. As a result, a classifier applied
all three evaluation measures. on such a feature space may not produce satisfying perfor-
We also observe that our method outperforms two stat-of- mance.
arts text representation learning methods: LSA and LDA.
6
https://radimrehurek.com/gensim/index.html 5. CONCLUSIONS
 [14] V. Nahar, X. Li, and C. Pang. An effective approach
In this paper, we have proposed a novel representation for cyberbullying detection. Communications in
learning method for cyberbullying detection, which is named Information Science and Management Engineering,
Embedding -enhanced Bag-of-Words. EBoW concatenates 2012.
BoW features, latent semantic features and bullying features [15] V. Nahar, X. Li, and C. Pang. An effective approach
together. Bullying features are derived based on word em- for cyberbullying detection. Communications in
beddings, which can capture the semantic information be- Information Science and Management Engineering,
hind words. When the final representation is learned, a lin- 3(5):238, 2013.
ear SVM is adopted to detect bullying messages. The ef- [16] H. C. Wu, R. W. P. Luk, K. F. Wong, and K. L.
fectiveness of our proposed model has been experimentally Kwok. Interpreting tf-idf term weights as making
verified through a real-world Twitter corpus. relevance decisions. ACM Transactions on
Information Systems (TOIS), 26(3):13, 2008.
6. REFERENCES [17] J.-M. Xu, K.-S. Jun, X. Zhu, and A. Bellmore.
Learning from bullying traces in social media. In
[1] D. M. Blei, A. Y. Ng, and M. I. Jordan. Latent Proceedings of the 2012 conference of the North
dirichlet allocation. the Journal of machine Learning American chapter of the association for computational
research, 3:993–1022, 2003. linguistics: Human language technologies, pages
[2] C.-C. Chang and C.-J. Lin. Libsvm: a library for 656–666. Association for Computational Linguistics,
support vector machines. ACM Transactions on 2012.
Intelligent Systems and Technology (TIST), 2(3):27, [18] M. Ybarra. Trends in technology-based sexual and
2011. non-sexual aggression over time and linkages to
[3] G. Chowdhury. Introduction to modern information nontechnology aggression. National Summit on
retrieval. Facet publishing, 2010. Interpersonal Violence and Abuse Across the Lifespan:
[4] K. Dinakar, R. Reichart, and H. Lieberman. Modeling Forging a Shared Agenda, 2010.
the detection of textual cyberbullying. In The Social [19] R. Zhao and K. Mao. Semi-random projection for
Mobile Web, 2011. dimensionality reduction and extreme learning
[5] M. Fekkes, F. I. Pijpers, A. M. Fredriks, T. Vogels, machine in high-dimensional space. Computational
and S. P. Verloove-Vanhorick. Do bullied children get Intelligence Magazine, IEEE, 10(3):30–41, 2015.
ill, or do ill children get bullied? a prospective cohort
study on the relationship between bullying and
health-related symptoms. Pediatrics,
117(5):1568–1574, 2006.
[6] G. Gini and T. Pozzoli. Association between bullying
and psychosomatic problems: A meta-analysis.
Pediatrics, 123(3):1059–1065, 2009.
[7] F. Godin, B. Vandersmissen, W. De Neve, and
R. Van de Walle. Named entity recognition for twitter
microposts using distributed word representations. In
Proceedings of the Workshop on Noisy User-generated
Text, pages 146–153, Beijing, China, July 2015.
Association for Computational Linguistics.
[8] J. Juvonen and E. F. Gross. Extending the school
grounds?âĂŤbullying experiences in cyberspace.
Journal of School health, 78(9):496–505, 2008.
[9] R. M. Kowalski, G. W. Giumetti, A. N. Schroeder,
and M. R. Lattanner. Bullying in the digital age: A
critical review and meta-analysis of cyberbullying
research among youth. 2014.
[10] T. K. Landauer, P. W. Foltz, and D. Laham. An
introduction to latent semantic analysis. Discourse
processes, 25(2-3):259–284, 1998.
[11] B. L. McLaughlin, A. A. Braga, C. V. Petrie, M. H.
Moore, et al. Deadly Lessons:: Understanding Lethal
School Violence. National Academies Press, 2002.
[12] T. Mikolov, K. Chen, G. Corrado, and J. Dean.
Efficient estimation of word representations in vector
space. arXiv preprint arXiv:1301.3781, 2013.
[13] T. Mikolov, I. Sutskever, K. Chen, G. S. Corrado, and
J. Dean. Distributed representations of words and
phrases and their compositionality. In Advances in
neural information processing systems, pages
3111–3119, 2013.