Authorship Analysis of Online Predatory Conversations using Character Level Convolution Neural Networks

Full text

Authorship Analysis of Online Predatory Conversations using
Character Level Convolution Neural Networks
Kanishka Misra, Hemanth Devarapalli, Tatiana R. Ringenberg, and Julia Taylor Rayz
Department of Computer and Information Technology, Purdue University, USA
Abstract— Authorship Attribution (AA) of written content
presents several advantages within the digital forensics domain.
While AA has been traditionally applied to long documents,
recent works have shown improved performance of neural AA
models on short texts such as tweets and online conversations
Concurrently, the rise of social media as well as a plethora of
chat messaging platforms have made it easier for teenagers to
be vulnerable to online predators. In this work, we present an
authorship attribution model that trains on a corpus of online
conversations involving predators, and perform subsequent
analysis of the message representations. Our results show
comparable performance relative to prior work for Authorship
Attribution and highlight differences between predatory and
non-predatory message styles.
I. INTRODUCTION
Authorship Attribution (AA) is a task of assigning an
author to a text whose author is unknown. An author is
selected from a set of candidate authors, given their known
texts as samples to learn their style [26, 10]. This task has
been successful in attributing authors of text samples at the
document level, such as blog posts [15], fan-fiction stories
[11], written pieces of literary text [16], books [17], etc.
More recently, with the advent of social media and the age
of short texts, such as text messages and tweets, research has
ventured into the authorship analysis of texts at the sentence
or phrase level [23, 25].
Apart from having direct applications within the realm of
digital forensics and plagiarism detection, authorship analysis
models have also been found to extract useful information
regarding the authors themselves, such as age or gender
[13]. By utilizing this feature of authorship analysis models,
we examine their application in the domain of online chat
conversations, specifically those involving online predators,
and present an analysis of messages that have been posted
by predators and non-predators as encoded by the models.
In this paper, we present a modified application of Author-
ship Attribution models, specifically those that are trained to
learn the writing styles of predatory users from their mes-
sages as found in online chat conversations. To that end, we
exploit recently proposed Neural Network models that have
been shown to perform well in Authorship Attribution of
short texts [21, 25] and analyze the difference in the encoded
representations of predatory and non-predatory messages.
The research setting of this paper consists of several
conversations, partly of predatory nature, where a predator
talks to a decoy who is pretending to be a minor. In prior
works, tasks focusing on a similar setting have aimed at (1)
analyzing the linguistic properties of predators, decoys and
real victims [8, 4]; (2) predicting whether a given user is
a predator or a non-predator[9, 2, 6, 5]; (3) Identifying the
specific lines that may contain predatory content [9]; and
(4) differentiating between offenders that wish to physically
meet their victims vs. keep the interaction to an online setting
[20]. The work presented here deviates from predicting
the user type, or word-choice based analysis of predatory
users, and focuses on analyzing the difference between the
messages as encoded by authorship attribution models.
Briefly, recent work in the authorship attribution of short
texts have shown convolution neural networks (CNN) trained
on character level units to outperform the state of the art
methods [21, 25]. Building on previous work, in this paper,
we propose modified variants of the CNN models: (1) AA-
CNN, that utilizes unigram and bigram vector representations
of characters to attribute each individual message to its
author; and (1) AA-CNN, that utilizes unigram and bigram
vector representations of characters to attribute each individ-
ual message to its author; and (2) AA-CNN-PC, that jointly
learns the objective of (1) but also contains an auxiliary layer
to learn whether the user is a predator or not, using the same
weights that encode the message style.
Specifically, we are trying to address the following re-
search questions:
1) Whether the newly proposed models, that learn from
multiple character n-gram signals produces comparable
performance compared to its predecessors.
2) Whether the encoded representations of the messages
from a predator differ from those by non-predators,
i.e., whether the AA-CNN model implicitly learns
predatory style along with author style, or does it
have to be jointly trained with explicit predatory signal
(AA-CNN-PC).
The contributions of this work are summarized as follows:
•We present two models for authorship attribution of
short texts, one that encodes style of individual mes-
sages using a combination of features inspired from
previous work; and one that jointly learns style as well
as the type of the author (predatory vs non-predatory).
•We present an analysis to test whether style based CNN
trained on messages encode the differences between the
various types of authors in the conversations (predatory
vs non-predatory).
The rest of the paper is organized as follows: Section
II introduces prior work relevant to the topic. Section III
describes our corpus, Section IV describes the model as well
2019 IEEE International Conference on Systems, Man and Cybernetics (SMC)
Bari, Italy. October 6-9, 2019
978-1-7281-4569-3/19/$31.00 ©2019 IEEE 623

as the baselines established in prior. Then, in Section V, we
present the results of our models compared to the baselines
and test whether our model implicitly learns to differentiate
between predators and non-predators (RQ2), we compare the
message representations from AA-CNN and AA-CNN-PC by
measuring the difference between encoded messages from
predators and non-predators and assess these differences
using statistical tests.
II. BACKGROUND
A. Authorship Analysis
A typical approach that has produced promising results
for Authorship Attribution is the use of stylistic features
such as character n-grams and function words [26, 1], or
syntactic features such as part of speech tags [14], along
with a machine learning algorithm, such as a Support Vector
Machines classifier. While several variations of features as
well as machine learning classifiers have been successful, this
success has mostly been in long documents. Performance of
traditional Authorship Attribution methods for short texts has
been shown to decline as the text length grows smaller [7].
B. Authorship Attribution of Short Texts
To circumvent this problem, several solutions have been
proposed. A simple workaround for making a text artificially
longer is treating several short texts from the same author
as one long one. Merging short texts, such as tweets, into
longer documents [3, 19] results in higher performance gains.
Schwartz et al. [23] introduce the concept of k-signatures,
which builds the unique profile of the author by taking all
the features that appear in the authors texts at least k%
of the time while not appearing in any other authors texts
during training. They achieve state of the art results in their
extensive experiments on tweets, a popular category of short
texts.Inches et al. [9] present a measurement of a users
unique profile by measuring the Kullback-Leibler Divergence
between each users vocabulary and the vocabulary of their
interlocutors in the chat. The metric is used to differentiate
between users that have conversations with more than one
other user.
While these methods achieve favorable results in the au-
thorship attribution of short text documents, they still require
manual feature selection (as in the case of k-signatures and
user profiles), or concatenation, all of which are done by
taking each user’s entire known texts into account.
To perform authorship attribution on shot texts by au-
tomatically and efficiently learning feature representations,
recent work has relied on using Neural Network models.
Specifically, Sari et al. [22] represent a document by a
continuous bag of character ngrams and learn dense repre-
sentations for the character ngrams which are then given as
input into a linear classifier with a softmax layer to achieve
improved performance in Authorship Attribution for two out
of five different corpora. Ruder et al [21] replicate the multi-
channel convolution neural network [12] and experimented
on a variety of web-scale data, including tweets and achieve
state of the art performance at the time by using a character-
level non-static channel (0.87 F1 score on tweets with
50 authors). Shrestha et al. [25] propose two very large
character level CNN models (large in terms of filter size) that
extract unigram and bigram level information, respectively,
from tweets and outperform word level CNNs, SVM based
models, and LSTMs in terms of authorship attribution (0.76
F1 score on tweets with 50 authors, separate corpus from
Ruder et al). We use Ruder et al. [21] and Shrestha et al [25]
as our baselines. All these models compared their results with
traditional methods such as SVMs over word and character
n-grams and reported improvements over them.
C. Conversations Involving a Predator
Within the domain of chat conversations that involve
predator, extensive research has been done in order to analyze
important linguistic differences between predators, teenagers,
and decoys [8, 4]. Machine learning models have been
trained on conversations involving a predator and a decoy
(pretending to be a minor) to detect whether a given user
is a predator or not. This was the major focus of the PAN-
2012 competition [9], where conversations with predators
and non-predators were combined to form a large corpus
of 2 million messages and participating teams were asked
to detect the predatory users. While this was successfully
solved by using a binary classifier, the second task, that of
detecting specific lines containing explicit predatory content
led to poor results, overall. The best performing model for the
second task used all the predicted predatory lines, resulting a
massive loss of Precision (0.09) but a high recall score (.89).
We hypothesize that this is because the traditional models
were unable to extract useful signals at the line level due to
the sparsity of their feature matrix. Finally, Ringenberg et al.
[20], and Seigfried-Spellar et al. [24] differentiated between
contact and fantasy offenders using a chat conversations from
Perverted Justice, an online repository of chat conversations
between predators and volunteer decoys.
The present work deviates from the rest of the studies
conducted on predatory conversations, since we attempt to
encode predator and non-predatory style in terms of an
authorship attribution model. Specifically, we probe these
models in their ability to not only differentiate between
authors, but also between the type of authors, i.e., whether
the author is a predator or not in terms of their style.
III. CORPUS
The corpus used in this work is built by combining
conversations from Perverted Justice (PJ), which hosts 623
chat logs of several volunteer decoys and their conversations
with online predators, and the PAN-2012 Sexual Predator
Identification corpus [9], which consists of messages involv-
ing a predator (crawled from PJ), and several conversations
from IRC chat logs. We combine all predator and decoy
authors from PJ with all the non-predatory conversations
from PAN, with restrictions that the message length in
consideration should be between 3 and 200 words, as well as
that an author should have at least 600 messages across all
624

Fig. 1: Proposed Neural Network Architectures, AA-CNN (left) and AA-CNN-PC (right)
the conversations that they take part in. This leaves us with
345 total authors. We randomly sample sets of 10 authors
and 50 authors to remain consistent with the analysis in [21,
25]. All html, email or URL strings were discarded from the
messages.
IV. METHODOLOGY
We propose two models, AA-CNN and AA-CNN-PC,
both of which are Convolution Neural Networks (CNN)
that accept character unigram and bigram input features to
encode each message’s representation. Then, depending on
the task (Authorship Attribution or Predator Classification),
the encoded vector is passed on to fully connected layers to
produce the output. Fig 1. illustrates the architectures of both
these models.
A. Continuous Character n-gram Embeddings
Character based features have been shown to be extremely
robust in Authorship Attribution models [26]. Traditionally,
these have been used as a discrete bag-of-feature repre-
sentations, while more recent work has found continuous
representations of characters and character n-grams, to have
performance gains in Authorship Attribution [22, 25]. We
choose to use character unigram and character bigrams for
our model, inpired by a combination of Ruder et al. [21]
and Shrestha et al. [25]. The convolution over unigrams will
produce short length windows (as used by Shrestha et al.
[25]), and over bigrams will use longer windows (as used
by Ruder et al. [21]) to encode the full message.
B. Character Level Convolution Neural Network
The input to the CNN is a text that is padded to a fixed
character length (the maximum length across the corpus),
n, represented by xi:n, where xi∈R1×kis the kdimen-
sional vector of the ith character-ngram in the text (either a
unigram, or a bigram).
The convolution layer slides over the input text with
different window sizes, h= [h1, h2, ..hl], and applies filters
weights, w∈R1×hi.k for each window size hi. Each filter
generates a new feature ciusing the following operation:
ci=g(w.xi:i+h−1+b)(1)
Where g(.)is a non linearity, in our case being a ReLU
operation, and b∈Ris the bias term present in each of
the layers. After a convolution over all the windows, the
features are concatenated to form a feature map for each
filter: cfilter = [c1;c2;c3,;..cn−hi+1]
Using the max-over-time pooling function, all mfeature
maps are reduced and concatenated to a vector p∈R1×m.
The pooled vectors for all lwindows are concatenated to get
the final output of the convolutional layers: s∈R1×lm. Since
we use two such convolution layers - one each for unigram
and bigram, we concatenate the final output of both layers
to get a fixed size vector v∈R1×2lm. This fixed size vector
is now given as input to a softmax layer for classifying the
author of the input text. The model is trained to minimize
the cross-entropy loss.
C. Auxiliary Layer for Predator Classification
To extract style based information about predatory users,
while also maintaining performance of the Authorship At-
tribution model, we adjust the AA-CNN model and add an
auxiliary layer to serve as a classifier for detecting predators,
resulting in the AA-CNN-PC model. Since the auxiliary layer
is jointly trained with the main Authorship Attribution layer,
the final loss for the AA-CNN-PC model is defined as:
Lfinal =LAA +LPC (2)
where LAA is the loss for the authorship attribution layer
(same as in AA-CNN) and LP C is the loss for the auxiliary
Predator Classification layer (binary cross entropy loss).
While training, we want the model to attribute the input
message to its author while also explicitly learning about
whether the author is a predator or not, as opposed to
AA-CNN, where if the model does learn about the author
being a predator or not, it will be implicit.
D. Model Training
For our models, we set the embedding dimension k= 100,
three windows h= [3,4,5], and m= 100 filters. This yields
a final vector v∈R1×600. The max length, nis inferred
from corpus and is found to be 200. We apply a dropout
of probability 0.5after the embedding layer, and train over
625

mini-batches of size 32 for 50 epochs with Adam as the
optimizer with a learning rate of 0.001.
E. Baseline Models
For comparison, we select two network architectures
which reported state of the art results in Authorship At-
tribution. The papers were selected based on how the text
was processed, which can be the whole document at once,
or sentence at a time. Our proposed network architecture
works sentence at a time, and hence we chose papers which
process text similarly. The selected papers work with short
texts, tweets in particular.
Ruder et al.[21] proposed character level CNNs for Au-
thorship Attribution that achieved the best results on tweets,
emails and reddit comments with 10 and 50 authors. They
use a network that processes single characters from the
input that are randomly initialized with a 300-dimensional
embedding layer. They use window sizes of 6,7, and 8 and
a feature map size of 100. They apply a dropout of 0.5 and
train their network for 15 epochs, with a Stochastic Gradient
Descent with the Adadelta update rule and a learning rate of
0.001.
Shrestha et al. [25] proposed a variant of Ruder’s [21]
architecture which achieved state of the art on tweets with
50 authors. The proposed architecture work with bigrams
instead of single character level embeddings. In their case,
the embedding size is 300, window sizes are 3, 4, and 5,
and the feature map size is 500. They use a dropout of 0.25
probability and train their network for 100 epochs using the
Adam optimizer with a learning rate of 0.0001.
V. EXPERIMENTS
We experiment on two sets of data: (1) A set of 10
randomly selected authors, with 5 being predators and 5
being non-predators; and (2) A set of 50 randomly selected
authors, with 25 being predators and 25 being non-predators.
We split our corpus into a set of 400, 100, and 100 sample
messages for each author as the training, development, and
testing set. Finally, we train our models and propose the
following two experiments.
A. Performance of AA-CNN and AA-CNN-PC models
The performance of our two models is compared to the
baselines - described in the previous section. We train the
baselines as stated in the papers that introduced them and
present results on the 10 and 50 author sets in Table 1. We
use the micro-averaged F1score as our comparison metric.
Some differences between differences between each of the
models in terms of the architectures (the feature embedding
size, as well as the feature map size) are highlighted in Table
1 as well.
1) Results: The results from Table 1 indicate comparable
performance of our models as compared to the baselines.
For the 10 authors set, both our models resulted in a slightly
lower F1 score than one of the baselines and a higher F1
score than another. For the 50 authors set, our AA-CNN-
PC model outperforms both baselines. Interestingly, the AA-
CNN shows improved performance for 10 authors compared
TABLE I: MICRO-AVERA GE D F1SC ORE S OF AA-CNN A ND
AA-CNN-PC CO MPARE D TO T HE B AS EL INE S FO R 10 A N D 50
AUT HO R SE T S
Model Model Architecture 10 Authors 50 Authors
Ruder et al., 2016 Emb.
size
300 0.5250 0.3524
Feature
maps
100
Shrestha et al., 2017 Emb.
size
300 0.5880 0.4474
Feature
maps
500
Ours (AA-CNN) Emb.
size
100×20.5570 0.4382
Feature
maps
100
Ours (AA-CNN-PC) Emb.
size
100×20.5490 0.4484
Feature
maps
100
to the AA-CNN-PC model, while for the 50-author set,
the opposite happens. Both models show that the difference
between the performance of 10 and 50 authors is smaller
than that of the baselines.
2) Discussion: The results from comparisons to our base-
lines indicate similar performance of our models to the state
of the art in Authorship Attribution of short texts. While we
are unable to outperform one of the baselines in the 10-author
set, we do outperform them in the 50-author set, but by a
very low margin. However, our model architecture is much
simpler than that of Shrestha et al. [25] we use a third of
the embedding size they used and a fifth of the feature map
size as compared to them. These parameters were set so that
we can easily add an auxiliary layer in AA-CNN-PC without
altering the batch-size and the number of epochs in order to
train at the same speed, without any memory issues.
Within our models, we find AA-CNN to beat the jointly
trained AA-CNN-PC model for the 10-author set, but the
opposite happens in the 50-author set experiments. We
hypothesize that the AA-CNN-PC model benefits from ob-
serving differences between predators and non-predators in
the 50-author set, since it trains on 5 times more examples
than in the case of the 10-author set, where the difference is
less pronounced. Due to the lack of sufficient examples, the
AA-CNN-PC might be overfitting in the predator classifica-
tion task, and in turn experiencing a slight increase in the
Authorship Attribution loss in the 10-author set. This effect
would be less when it gets more examples to train on. We
leave this exploration of joint-training loss as something to
be covered in future work.
B. Probing message representations for differences between
Predators and non-Predators
Both our proposed models learn to encode chat messages
using their character unigram and bigram embeddings and
use this representation to attribute each message to its author.
While the AA-CNN only serves as an authorship attribution
626

Fig. 2: t-SNE visualization of message representations in the training set, extracted AA-CNN model (left) and AA-CNN-PC
(right). Four example messages have been shown, two each from the predator and non-predator groups.
model, the AA-CNN-PC model jointly trains to encode
author style as well as the type of author (predator vs.
non-predator) using its auxiliary binary classification layer.
In this section, we analyze the differences in the encoded
message representations in both the models. Specifically,
we are interested in probing the message representations
in order to learn whether the AA-CNN model implicitly
extracts signals for differences between predators and non-
predators, as compared to the AA-CNN-PC model (which
we expect, learns to do this explicitly, due to the additional
layer). In order to compare message representations in both
models, we first extract the final representations from the
CNN layers of the trained models and separate them into
two groups (predator vs non-predator) using the ground truth
labels. Then, in order to measure the differences between the
representations of the groups, we propose a metric, the Mean
Average Similarity (MAS), which in our case, will measure
the similarity of predatory messages to either other predatory
messages, or to non-predatory messages. Mathematically, for
message vectors va
iand vb
j, belonging to group A(of length
Ni) and B(of length Nj) respectively, the MAS is given
as:
MAS(va
i, vb
j) = 1
Ni
1
Nj
Ni
X
i
Nj
X
j
cos(va
i, vb
j)(3)
The difference between the two groups in our case is further
measured as the change in the MAS (∆MAS)between
the message representations of predators and that of non-
predators. It is computed as:
∆MAS =M AS (vpredator
i, vpredator
j)1i6=j−
MAS(vpredator
i, vnon−predator
j)(4)
This metric measures the difference between the similarity of
predatory messages to each other compared to non-predatory
messages. For 10000 iterations, we randomly sample 1000
predatory and 1000 non-predatory messages and compute
the ∆MAS for each iteration. Then, we conduct a one-
sided t-test to measure the significance of MAS for both
the AA-CNN, as well as AA-CNN-PC models. The results
are summarized in Table II.
TABLE II: CHA NG E IN T HE MEA N AVER AGE SIMILARITY
(PRE DATO R - NO N PRE DATO R)I N AA-CNN AS W E LL A S
AA-CNN-PC FO R 2000 RE SA M PL ES OV E R 10000 ITERATIONS.
Model ∆MAS
AA-CNN 0.021 (t= 1048.3,p= 2.2×10−16)
AA-CNN-PC 0.025 (t= 1285.8,p= 2.2×10−16)
1) Results: The results shown in Table II indicate a
statistically significant difference between the similarity of
predator messages with other predatory messages, as com-
pared to with non-predatory messages. This difference is
greater in AA- CNN-PC as compared to the AA-CNN model.
2) Visualizing the encoded representations of messages:
To visually highlight the differences between predatory and
the non-predatory messages, we computed a t-SNE [18]
visualization with default parameters for both the AA-CNN
and the AA-CNN-PC model. The plots are shown in Fig. 2.
3) Discussion: From the results of this experiments, we
find that both models learn the differences between messages
from a predator and those from a non-predator. This is seen
both in our statistical comparisons, as well as in the t-SNE
projections (Fig. 2). While the AA-CNN model is never
explicitly provided an information about whether the user
of the message is a predator or not (as is in the case of AA-
CNN-PC), it still encodes the differences, although not as
much as in the model that jointly learns these differences.
The AA-CNN-PC jointly adapts to encoding styles of the
kind of author while performing the task of authorship
attribution, and thus it is not surprising that the message
627

from predators were more closer together in this model as
compared to the simple AA-CNN model.
VI. CONCLUSION
In this paper, we presented two models, one that is trained
to do authorship attribution on a corpus of chat messages
from predators and non-predators, and one that performs
authorship attribution and jointly learns whether the author of
the message is a predator or a non-predator. Our models were
simpler yet comparable in terms of performance with our
baselines. We also presented an analysis of the messages as
encoded by our CNN models and found that the Authorship
Attribution model was able to implicitly encode information
about the kind of author without any explicit signal, and
differentiated between predators and non-predators, while
this difference was more pronounced in the model that
explicitly received signal about the author type.
Predators present a risk to minors online and one way
of reducing this risk is to develop tools to detect predatory
behavior from evidence, usually found in the form of chat
conversations. While the models in this work did show
certain aspects of encoding predatory behavior purely from
style (characters), in the future, we would like to work
on combining this with more contextual models, such as
sequence encoders. Using the context of a chat messages,
we hope to build better models of detecting risk over the
internet to aid in making it a better space for everyone.
ACK NO WL E DG EME NTS
This research was partially supported by Purdue Research
Foundation.
REFERENCES
[1] S. Argamon et al. “Stylistic text classification using func-
tional lexical features”. In: Journal of the American Soci-
ety for Information Science and Technology 58.6 (2007),
pp. 802–822.
[2] D. Bogdanova, P. Rosso, and T. Solorio. “On the impact
of sentiment and emotion based features in detecting online
sexual predators”. In: Proceedings of the 3rd workshop
in computational approaches to subjectivity and sentiment
analysis. Association for Computational Linguistics. 2012,
pp. 110–118.
[3] S. R. Boutwell. “Authorship attribution of short messages us-
ing multimodal features”. PhD thesis. Monterey, California.
Naval Postgraduate School, 2011.
[4] M. M. Chiu, K. C. Seigfried-Spellar, and T. R. Ringenberg.
“Exploring detection of contact vs. fantasy online sexual
offenders in chats with minors: Statistical discourse analysis
of self-disclosure and emotion words”. In: Child abuse &
neglect 81 (2018), pp. 128–138.
[5] M. Ebrahimi, C. Y. Suen, and O. Ormandjieva. “Detecting
predatory conversations in social media by deep convolu-
tional neural networks”. In: Digital Investigation 18 (2016),
pp. 33–49.
[6] M. Ebrahimi et al. “Recognizing predatory chat documents
using semi-supervised anomaly detection”. In: Electronic
Imaging 2016.17 (2016), pp. 1–9.
[7] M. Eder. “Does size matter? Authorship attribution, small
samples, big problem”. In: Digital Scholarship in the Hu-
manities 30.2 (Nov. 2013), pp. 167–182. ISS N: 2055-7671.
[8] K. Guice. Predators, decoys, and teens: A corpus analysis
of online language. Hofstra University, 2016.
[9] G. Inches and F. Crestani. “Overview of the International
Sexual Predator Identification Competition at PAN-2012.”
In: CLEF (Online working notes/labs/workshop). Vol. 30.
2012.
[10] P. Juola et al. “Authorship attribution”. In: Foundations and
Trends R
in Information Retrieval 1.3 (2008), pp. 233–334.
[11] M. Kestemont et al. “Overview of the author identification
task at PAN-2018: cross-domain authorship attribution and
style change detection”. In: Working Notes Papers of the
CLEF 2018 Evaluation Labs. Avignon, France, September
10-14, 2018/Cappellato, Linda [edit.]; et al. 2018, pp. 1–25.
[12] Y. Kim. “Convolutional neural networks for sentence classi-
fication”. In: arXiv preprint arXiv:1408.5882 (2014).
[13] M. Koppel, S. Argamon, and A. R. Shimoni. “Automatically
categorizing written texts by author gender”. In: Literary and
linguistic computing 17.4 (2002), pp. 401–412.
[14] M. Koppel and J. Schler. “Exploiting stylistic idiosyncrasies
for authorship attribution”. In: Proceedings of IJCAI’03
Workshop on Computational Approaches to Style Analysis
and Synthesis. Vol. 69. 2003, pp. 72–80.
[15] M. Koppel, J. Schler, and S. Argamon. “Authorship attribu-
tion in the wild”. In: Language Resources and Evaluation
45.1 (2011), pp. 83–94.
[16] M. Koppel, J. Schler, and E. Bonchek-Dokow. “Measur-
ing differentiability: Unmasking pseudonymous authors”.
In: Journal of Machine Learning Research 8.Jun (2007),
pp. 1261–1276.
[17] M. Koppel et al. “Unsupervised decomposition of a doc-
ument into authorial components”. In: Proceedings of the
49th Annual Meeting of the Association for Computational
Linguistics: Human Language Technologies-Volume 1. Asso-
ciation for Computational Linguistics. 2011, pp. 1356–1364.
[18] L. v. d. Maaten and G. Hinton. “Visualizing data using
t-SNE”. In: Journal of machine learning research 9.Nov
(2008), pp. 2579–2605.
[19] G. K. Mikros and K. Perifanos. “Authorship attribution in
greek tweets using author’s multilevel n-gram profiles”. In:
2013 AAAI Spring Symposium Series. 2013.
[20] Tatiana Ringenberg et al. “Exploring Automatic Identifica-
tion of Fantasy-Driven and Contact-Driven Sexual Solic-
itors”. In: 2019 Third IEEE International Conference on
Robotic Computing (IRC). IEEE. 2019, pp. 532–537.
[21] S. Ruder, P. Ghaffari, and J. G. Breslin. “Character-level and
multi-channel convolutional neural networks for large-scale
authorship attribution”. In: arXiv preprint arXiv:1609.06686
(2016).
[22] Y. Sari, A. Vlachos, and M. Stevenson. “Continuous n-gram
representations for authorship attribution”. In: Proceedings
of the 15th Conference of the European Chapter of the
Association for Computational Linguistics: Volume 2, Short
Papers. Vol. 2. 2017, pp. 267–273.
[23] R. Schwartz et al. “Authorship attribution of micro-
messages”. In: Proceedings of the 2013 Conference on
Empirical Methods in Natural Language Processing. 2013,
pp. 1880–1891.
[24] K. C. Seigfried-Spellar et al. “Chat Analysis Triage Tool:
Differentiating Contact-Driven vs. Fantasy-Driven Child Sex
Offenders”. In: Forensic science international (2019).
[25] P. Shrestha et al. “Convolutional neural networks for author-
ship attribution of short texts”. In: Proceedings of the 15th
Conference of the European Chapter of the Association for
Computational Linguistics: Volume 2, Short Papers. Vol. 2.
2017, pp. 669–674.
[26] E. Stamatatos. “A survey of modern authorship attribution
methods”. In: Journal of the American Society for informa-
tion Science and Technology 60.3 (2009), pp. 538–556.
628