=Paper=
{{Paper
|id=Vol-2125/paper_105
|storemode=property
|title=Multi-Language Neural Network Model with Advance Preprocessor for Gender Classification over Social Media: Notebook for PAN at CLEF 2018
|pdfUrl=https://ceur-ws.org/Vol-2125/paper_105.pdf
|volume=Vol-2125
|authors=Kashyap Raiyani,Teresa Gonçalves,Paulo Quaresma,Vítor Beires Nogueira
|dblpUrl=https://dblp.org/rec/conf/clef/RaiyaniGQN18
}}
==Multi-Language Neural Network Model with Advance Preprocessor for Gender Classification over Social Media: Notebook for PAN at CLEF 2018==
https://ceur-ws.org/Vol-2125/paper_105.pdf
Multi-Language Neural Network Model with Advance
Preprocessor for Gender Classification over Social
Media
Notebook for PAN at CLEF 2018
Kashyap Raiyani, Teresa Gonçalves, Paulo Quaresma and Vitor Beires Nogueira
Computer Science Department, University of Évora, Portugal
{kshyp,tcg,pq,vbn}@uevora.pt
Abstract This paper describes approaches for the Author Profiling Shared Task
at PAN 2018. The goal was to classify the gender of a Twitter user solely by their
tweets. Paper explores a simple and efficient Multi-Language model for gender
classification. The approach consists of tweet preprocessing, text representation
and classification model construction. The model achieved the best results on
the English language with an accuracy of 72.79%; for the Spanish and Arabic
languages the accuracy was 72.20% and 64.36%, respectively.
Keywords: Author Profiling, Gender Classification, Twitter, PAN
1 Introduction
The Author Profiling is the identification of demographic features of an author by exam-
ining his written text. Recently, it has attracted the attention of research community due
to its potential applications in forensic, security, marketing, fake profiles identification
on online social networking sites, capturing sender of harassing messages etc. Knowing
the profile of an author could be of key importance; for instance, from a forensic lin-
guistics perspective, being able to know what is the linguistic profile of a suspected text
message (language used by a certain type of people) and identify characteristics (lan-
guage as evidence) just by analyzing the text would certainly help considering suspects.
Following this ideology, PAN [26] is a series of scientific events and shared tasks on
digital text forensics and stylometry. In PAN 2018 the focus was on Author Identifica-
tion, Author Profiling, and Author Obfuscation. The Author Identification [11] task was
to identify who wrote the document given a document. The subtasks focus on crossdo-
main authorship attribution applied in fanfiction and on style change detection. Where
in Author Profiling [21], the task was to find author’s gender whereas text and image
might be used as information sources of tweets in English, Spanish and Arabic. Lastly,
Author Obfuscation [17], given the document, the tasks were author masking and ob-
fuscation evaluation. Basically, this task works against identification and profiling by
automatically paraphrasing a text to obfuscate its author’s style.
This paper emphasizes the author profiling shared-task and delivers brief about the
recent related work towards gender identification in social media and the technology
and methods used in building systems. Later on, it also shows the shared-task submit-
ted system modeling and its results. It is organized in the following manner: section
2 introduces past research over gender classification and presents a brief introduction
to gender classification and its components; section 3 outlines the stylometric data and
preprocessing; section 4 describes different data representations; section 5 describes
methods and system modeling; section 6 discusses the experimental results obtained.
Conclusion and future work are highlighted in Section 7.
2 Related Work
In recent years, efforts have been made by the research community to develop bench-
mark corpora for author profiling task. The most prominent effort in this regard is the
series of PAN competitions on author profiling ( [18–23]). The focus of author profil-
ing has settled much on the social media, including languages other than English. This
section will cover the previous findings on the author profiling.
The work from Basile et. al. [2] reported having the best performance at the PAN
2017 [22]. Their system used a linear support vector machine (SVM) with word un-
igrams and character 3- to 5-grams as features. Apart from that, they concluded that
additional features, including part of speech tags, additional datasets, geographic enti-
ties and Twitter handles hurt rather than improve performance.
The approach in Matej et. al. [14] has mainly consisted of tweet preprocessing,
feature construction, feature weighting and classification model construction with lin-
ear SVM, logistic regression, logistic regression bagging, random forest, and XGBoost.
The best results were obtained with Logistic Regression and the Bagging & Voting did
not improve the results. Here, the main features used were different types of character
and word n-grams with supporting features like POS tag sequences, emoji counts, char-
acter flood counts, language variety specific word lists and document sentiments. They
concluded with that the most difficult part of the task was finding the right features and
properly weighing their combination. This approach was able to achieve 2nd highest
results among author profiling task at PAN 2017 [22].
Eric et. al. [27] used the MicroTC (µTC) framework tool for classification. It follows
a simple approach to text classification and converts the problem of text classification
to a model selection problem using several simple text transformations, a combination
of tokenizers, a term-weighting scheme, and finally, it classifies using a Support Vector
Machine. In particular, their system works best when Hill Climbing [7] procedure is
applied over the best configuration found by a Random Search [8]. Here, the main idea
behind Hill Climbing is to explore the configuration’s neighborhood and greedily move
to the best neighbor.
The work from Vollenbroek et. al. [4] reported having the best performance at the
PAN 2016 [23]. They trained SVM linear model using the feature traits like N-grams,
starts a capital sentence, capitalized tokens, capital letters, endings with punctuation,
punctuation by sentence, average word length and average sentence length, out of dic-
tionary words, vocabulary richness, function words, parts of speech, and emotions.
Pashutan et. al. [16] states the extraction of stylistic and lexical features for training
a logistic regression model. Here, lexical features such as unigrams and bigrams were
used. Their approach stood first place for gender detection in English and tied for second
place in terms of joint accuracy in author profiling PAN 2016 [23].
There are several interesting works on author profiling from the perspective of the
common theoretical framework which involves several disciplines such as psychology,
(computational) linguistics or even neurology. On the language perspective, features
like word unigrams and bigrams, character tetragrams, average spelling error and part
of speech (POS) n-grams are considered. In concerned with machine learning algo-
rithms, logistic regression, support vector machine (SVC), multinomial NB, bernoulli
NB, ridge classifier, adaBoost classifier and deep learning systems are used. As de-
scribed, there are a lot of meta-variables that need to be chosen while designing a clas-
sification system: the features to be used as input and the machine learning algorithm
and its parameter values. Before proceeding with the methodology, some time was taken
to understand the data and the requirement of preprocessing.
3 Data Analysis & Preprocessing
The focus of this year’s author profile task [21] was on gender identification on Twitter,
where text and images were provided as information sources. The languages addressed
were English, Spanish, and Arabic. The provided training data set had 3000 Twitter
users labeled with gender and, for each user, a total of 100 tweets and 10 images were
provided. Authors were grouped by the language of their tweets. The provided data
had equal class distribution, meaning that 50% of tweets were from the male and rest
50% were from the female genders. The data had one XML file with 100 tweets for
each user. To retrieve the tweets from the XML files, python’s ElementTree XML API1
was used. This API has a parse feature which helps to translate XML tags to a tree
and different tags are considered as nodes. By traveling them through root node, users’
information was retrieved and stored in one CSV file; the corresponding gender was
retrieved from a separate text file. Initially, there were total 3000 XML files with 100
tweets each and after the above-mentioned process, a single CSV file with 300000 row
of tweets mapped with the corresponding gender was obtained.
The next subsections describe the text preprocessing steps and preprocessing done
for each dataset.
3.1 Preprocessing Steps
This subsections discuss the different preprocessing steps attained to build the input
features for the machine learning algorithm.
Merging of Stem Words. Words ending with ing, ious and ment were merged into
the corresponding root word. A regular expression was created and used to make this
transformation.
1
https://docs.python.org/2/library/xml.etree.elementtree.html
Character Repetition and Stopwords. The extra repetition of characters in a word
(for example "tooooo mucccch loveee isss nottt gooooood") were removed using the
regular expression followed by the removal of stopwords.
Users in Tweets. The main characteristic of tweets is to have ’@’ for mentioning user
and ’#’ for tagging. ’@’ is always followed by the name of the other user. But when you
consider the unique users in the provided dataset where 300000 tweets are present, the
question comes to the mind is this really necessary to include all the @usernames?
or should they be replaced by @user? with this in mind, @usernames were replaced
by word ’user’.
Symbolic Emoji Replacement. In day to day social media writing, people often use
the symbolic emojis. This step handles such emojies which are not in form of regular
emoji. Table 1 shows the replacements done.
Symbol Replaced with
:) : ) :-) (: ( : (-: :’) em_smile
:D : D :-D xD x-D XD X-D em_laugh
<3 :* em_love
;-) ;) ;-D ;D (; (-; em_wink
:-( : ( :( ): )-: em_sad
:,( :’( :"( em_cry
X( >:-( >:( X-( em_angry
Table 1. Symbolic Emoji Replacement
Segmentation. One challenge was to identify the meaning of the # or hashtag. There
are couple of tools available like nltk Segmentation [13] and Ekphrasis Segmenta-
tion [3] that can be used to segment and split hashtags.Here, Ekphrasis was used for
segmentation. Apart from that, another tool present in the Ekphrasis library was used to
replace a particular segment of the tweets with its corresponding type. The types con-
sidered were ’url’, ’email’, ’percent’, ’money’, ’phone’, ’time’, ’date’ and ’number’.
This library uses the regular expressions to normalize text segments.
Communication Abbreviations. Another preprocessing step was to remove day to
day communication abbreviations. Like consider two sentences "I am in love with you"
and "im in luv wid u". Both the sentences have the same meaning and might be written
by the same author but the machine will see them as two different representations. The
motivation is to reduce the confusion/possibility for the machine learning model. Table
2 presents the preprocessing done over a day to day communication abbreviations.
3.2 Data Preprocessing
For the Arabic dataset no preprocessing was done but for the English dataset, all the
above mentioned preprocessing approaches were performed before the text representa-
tion to be presented to the machine learning algorithm.
Word Replaced with
app, wil, im, al, sx, u application, will, i am, all, sex, you
r, y, hv, c, bcz or coz are, why, have, see, because
ppl or pepl, nd, hw people, and, how
Table 2. Abbreviation Preprocessor
One of the most common things that Spanish native speakers do when they text
is omitting some of the letters in words and using just the endings. For example, the
"es" sound at the beginning of words like "estoy", "estás" or "estamos" will be dropped,
leaving with "toy", "tás" and "tamos" instead. Dropping the "d" is widespread orally,
and it is also used when texting. For example, "cansado" is "cansao" and "todo" is
"to" or "too." If a friend texts "toy cansao" it means "estoy cansado" or "too tá listo" it
means "todo está listo." With this much versatility in the language, it was hard to design
the preprocessor without having Spanish language background. For that matter, only
symbolic emojis were replaced (see table 1) followed by the removal of the Spanish
stopwords. Here, merging of stem words was not done.
4 Text Representation
This section describes the text representation before passed to the machine learning
algorithm.
A sequential model can be designed for text classification where the text sequence is
fed to the machine learning algorithm2 . Nonetheless, these models only use numerical
data and therefore a conversion was needed which involves two subprocesses: Integer
Encoding and One-Hot Encoding [12].
Integer Encoding. Here, each unique word/text value is assigned an integer value
for example, "red" is 1, "green" is 2, and "blue" is 3. This process is known as label
encoding or an integer encoding and is easily reversible. The integer values have a
natural ordered relationship between each other and machine learning algorithms may
be able to understand and harness this relationship. For example, ordinal variables like
the "place" example above would be a good example where a label encoding would
be sufficient. But for sentences where no such ordinal relationship exists, the integer
encoding is not enough which leads to the next encoding One-Hot encoding.
red green blue
1 2 3
One-Hot Encoding. Here, the integer encoded variable is removed and a new binary
variable is added for each unique integer value. In the "color" variable example, there
2
https://keras.io/getting-started/sequential-model-guide/
are 3 words and therefore 3 binary variables are needed. For color, "1" value is placed
in the binary variable and "0" values for rest of the colors. For example:
red green blue
1 0 0
0 1 0
0 0 1
Representation. Both for English and Spanish, a word dictionary was created in which,
all the unique words were indexed; the index was then used as the word id. Using Integer
Encoding and One-Hot Encoding concepts, these ids were portrayed as integers first and
then as a binary matrix preserving the word order of the sentences.
For better understanding, assume the English dictionary {"country": 1, "very": 2,
"I": 3, "love": 4, "my": 5}. Considering the Figure 1, both sentences use the same word
set but their sentence ordering is preserved using the dictionary index and one-hot ap-
proach.
I love my country my country I love
3451 5134
[ [00100] [00010] [00001] [10000] ] [ [00001] [10000] [00100] [00010] ]
Figure 1. English Text Converstion
The given Arabic dataset was smaller in size compared to the English and Spanish
ones; only 1500 XML files were provided. Here, XML files were not trasnformed to a
CSV file; rather a single text file was created where each line contained a single tweet
from the user. These lines were prefixed with author’s gender. As a text representation,
a character unigram representation was used.
5 System Modeling
For the English and Spanish languages, a dense Neural Network was considered; for
Arabic, FastText [9] library was used. FastText is an open-source library developed by
Facebook Open Source for researchers to learn and test text classification problems. It
provides word vectors for 157 languages and supervised models for 8 datasets using a
character level n-gram approach. For research and development porpose, Keras [5] was
used as front-end and Tensorflow [1] as back-end.
The next sub-sections present the system architecture design.
5.1 Dense Neural Network
After experimenting with different methods and machine learning algorithms, a simple
Dense architecture was used to create the final model for submission. The figure 2
represents the architecture layers of the Neural Network.
After several iterations, the number of layers for the architecture was set to three.
The 1st hidden layer is having 1024 nodes which take the input from the input layer and
the 2nd layer had 512 nodes and the last layer had 256 nodes.
Inbetween each layers, mathamatical activation functions like Relu, Sigmoid and
Softmax were used. The ordering of this activation function had very high impact on
the results.
The table 3 presents the experiment parameters.
Figure 2. System Architecture
Parameter Value
Maximum Number of Words 2000
Validation Split Ratio 0.2
Epochs 2
Batch Size 32
Optimizer Adam
Table 3. Experiment Parameters
5.2 FastText
According to Joulin et. al. [10], a simple and efficient baseline for sentence classifi-
cation is to represent sentences as a bag of words (BoW) and train a linear classifier,
e.g., a logistic regression or an SVM [25, 28]. However, linear classifiers do not share
parameters among features and classes. For each example, y is a scalar that takes binary
values (ŷ ∈ [0, 1]), while x is a vector of length N (where N is the number of features).
Here, the the output is a linear combination of the input features (i.e. a weighted sum of
these features plus the biases).
P
N
ŷ = i (wi · xi ) + b or (w · x + b)
where x and w are vectors of length N and the product x·w produces a scalar. As can
be seen, a separate weight wi for each input feature xi is considered and these weights
are independent by all means. This shows that there is no parameter sharing among
features for linear classifiers. This possibly limits their generalization in the context of
large output space where some classes have very few examples. Common solutions to
this problem are to factorize the linear classifier into low-rank matrices [6, 15] or to use
multilayer neural networks [24, 29]
Figure 3 shows a simple linear model with rank constraint architecture which was
used for Arabic text classification (with the help of FastText).
Figure 3. Model architecture of FastText for a sentence with N ngram features x1 , . . . , xN . The
features are embedded and averaged to form the hidden variable
An Architecture similar to the cbow model of Mikolov et. al. [15], where the middle
word is replaced by a label, here, the first weight matrix is taken as a look-up table over
the words and then the word representations are averaged into a text representation,
which is in turn fed to a linear classifier. The text representation is a hidden variable
which can be potentially be reused. The softmax function was used to compute the
probability distribution over the predefined classes. For a set of N documents, this leads
to minimizing the negative loglikelihood over the classes
PN
− N1 n=1 yn log(f (BAxn ))
where xn is the normalized bag of features of the nth document, yn the label, A and B
the weight matrices. This model has trained asynchronously on multiple CPUs using
stochastic gradient descent and a linearly decaying learning rate.
The parameters of training were 150000 sentences as an input(or N ), learning rate
of 0.25, hidden value of 100(or h), 40 epochs and softmax loss function. With this, the
system was able to learn 99.98% on a train set.
6 Results and Discussion
Table 4 presents the results provided by the PAN 2018 organizing committee for the
systems described in the previous section.
Language Accuracy(%)
English 72.79
Spanish 64.36
Arabic 72.20
Table 4. Results for PAN 2018 Author Profiling Task
From the Table 4, one can say that the English and Arabic systems performed better
than the Spanish one. This difference can probably be explained by the fact that not
all described preprocessing filters steps were applied to the Spanish system: only the
simple regular expression transformation, stopwords removal and emoji normalization
were done. Apart from this, given the amount of dataset, with provided hyper parame-
ters, English & Spanish model reached to the saturation point of the designed model.
7 Conclusion and Future Work
This paper presents a system for multi-language author profiling task for gender classi-
fication and presents the findings from the related work. Further, it concludes that, for
the current author profiling task, a seemingly simple system using fully connected ar-
chitecture proves beyond average performance. Paper also described the preprocessing
techniques used, the methodology of approach and the conducted experiments.
As mentioned in the previous section 6, models reached its saturation point for the
design. This leads to the one limitation of this approach. Proposed systems are only
suitable when having a small dataset; with the large dataset, the dictionary index will
be huge and the system won’t be able to convert all the words to a binary matrix.
In this model words that are not found in the dictionary are omitted. As future
work, and to address this issue, a neighboring or similar word could be used; here
each word would be added to a list from which neighboring/similar word would be
founded. This will help in categorizing unseen words to the model. Thus, all the words
are included and there is no omission of words. Another line of future work can be to
include semantic meaning from the sentences, extracting ontological features either by
Part of Speech (POS) tagging or Entity Extraction (EE).
Acknowledgement
This research was supported by AGATHA (Intelligent analysis system of open infor-
mation sources for surveillance/crime control) project and by the University of Évora.
References
1. Abadi, M., Barham, P., Chen, J., Chen, Z., Davis, A., Dean, J., Devin, M., Ghemawat, S.,
Irving, G., Isard, M., Kudlur, M., Levenberg, J., Monga, R., Moore, S., Murray, D.G.,
Steiner, B., Tucker, P., Vasudevan, V., Warden, P., Wicke, M., Yu, Y., Zheng, X.:
Tensorflow: A system for large-scale machine learning. In: Proceedings of the 12th
USENIX Conference on Operating Systems Design and Implementation. pp. 265–283.
OSDI’16, USENIX Association, Berkeley, CA, USA (2016),
http://dl.acm.org/citation.cfm?id=3026877.3026899
2. Basile, A., Dwyer, G., Medvedeva, M., Rawee, J., Haagsma, H., Nissim, M.: N-GrAM:
New Groningen Author-profiling Model—Notebook for PAN at CLEF 2017. In: Cappellato
et al. [22], http://ceur-ws.org/Vol-1866/
3. Baziotis, C., Pelekis, N., Doulkeridis, C.: Datastories at semeval-2017 task 4: Deep lstm
with attention for message-level and topic-based sentiment analysis. In: Proceedings of the
11th International Workshop on Semantic Evaluation (SemEval-2017). pp. 747–754.
Association for Computational Linguistics, Vancouver, Canada (August 2017)
4. Busger op Vollenbroek, M., Carlotto, T., Kreutz, T., Medvedeva, M., Pool, C., Bjerva, J.,
Haagsma, H., Nissim, M.: GronUP: Groningen User Profiling—Notebook for PAN at
CLEF 2016. In: Balog et al. [23], http://ceur-ws.org/Vol-1609/
5. Chollet, F., et al.: Keras. https://keras.io (2015)
6. H., S.: Dimensions of meaning. In: Proceedings of the 1992 ACM/IEEE Conference on
Supercomputing. pp. 787–796. Supercomputing ’92, IEEE Computer Society Press, Los
Alamitos, CA, USA (1992),
http://dl.acm.org/citation.cfm?id=147877.148132
7. Jacobson, S.H., Yücesan, E.: Analyzing the performance of generalized hill climbing
algorithms. Journal of Heuristics 10(4), 387–405 (Jul 2004),
https://doi.org/10.1023/B:HEUR.0000034712.48917.a9
8. James, B., Yoshua, B.: Random search for hyper-parameter optimization. J. Mach. Learn.
Res. 13, 281–305 (Feb 2012),
http://dl.acm.org/citation.cfm?id=2188385.2188395
9. Joulin, A., Grave, E., Bojanowski, P., Mikolov, T.: Bag of tricks for efficient text
classification. arXiv preprint arXiv:1607.01759 (2016)
10. Joulin, A., Grave, E., Bojanowski, P., Mikolov, T.: Bag of tricks for efficient text
classification. CoRR abs/1607.01759 (2016),
http://arxiv.org/abs/1607.01759
11. Kestemont, M., Tschugnall, M., Stamatatos, E., Daelemans, W., Specht, G., Stein, B.,
Potthast, M.: Overview of the Author Identification Task at PAN-2018: Cross-domain
Authorship Attribution and Style Change Detection. In: Cappellato, L., Ferro, N., Nie, J.Y.,
Soulier, L. (eds.) Working Notes Papers of the CLEF 2018 Evaluation Labs. CEUR
Workshop Proceedings, CLEF and CEUR-WS.org (Sep 2018)
12. Lantz, B.: Machine Learning with R. Packt Publishing (2013)
13. Loper, E., Bird, S.: Nltk: The natural language toolkit. In: Proceedings of the ACL-02
Workshop on Effective Tools and Methodologies for Teaching Natural Language
Processing and Computational Linguistics - Volume 1. pp. 63–70. ETMTNLP ’02,
Association for Computational Linguistics, Stroudsburg, PA, USA (2002),
https://doi.org/10.3115/1118108.1118117
14. Martinc, M., Škrjanec, I., Zupan, K., Pollak, S.: PAN 2017: Author Profiling - Gender and
Language Variety Prediction—Notebook for PAN at CLEF 2017. In: Cappellato et al. [22],
http://ceur-ws.org/Vol-1866/
15. Mikolov, T., Chen, K., Corrado, G., Dean, J.: Efficient estimation of word representations in
vector space. CoRR abs/1301.3781 (2013), http://arxiv.org/abs/1301.3781
16. Modaresi, P., Liebeck, M., Conrad, S.: Exploring the Effects of Cross-Genre Machine
Learning for Author Profiling in PAN 2016—Notebook for PAN at CLEF 2016. In: Balog
et al. [23], http://ceur-ws.org/Vol-1609/
17. Potthast, M., Hagen, M., Schremmer, F., Stein, B.: Overview of the Author Obfuscation
Task at PAN 2018: A New Approach to Measuring Safety. In: Cappellato, L., Ferro, N.,
Nie, J.Y., Soulier, L. (eds.) Working Notes Papers of the CLEF 2018 Evaluation Labs.
CEUR Workshop Proceedings, CLEF and CEUR-WS.org (Sep 2018)
18. Rangel, F., Celli, F., Rosso, P., Potthast, M., Stein, B., Daelemans, W.: Overview of the 3rd
Author Profiling Task at PAN 2015. In: Cappellato et al. [18]
19. Rangel, F., Rosso, P., Chugur, I., Potthast, M., Trenkmann, M., Stein, B., Verhoeven, B.,
Daelemans, W.: Overview of the 2nd Author Profiling Task at PAN 2014. In: Cappellato
et al. [19]
20. Rangel, F., Rosso, P., Koppel, M., Stamatatos, E., Inches, G.: Overview of the Author
Profiling Task at PAN 2013. In: Forner et al. [20]
21. Rangel, F., Rosso, P., Montes-y-Gómez, M., Potthast, M., Stein, B.: Overview of the 6th
Author Profiling Task at PAN 2018: Multimodal Gender Identification in Twitter. In:
Cappellato, L., Ferro, N., Nie, J.Y., Soulier, L. (eds.) Working Notes Papers of the CLEF
2018 Evaluation Labs. CEUR Workshop Proceedings, CLEF and CEUR-WS.org (Sep
2018)
22. Rangel Pardo, F., Rosso, P., Potthast, M., Stein, B.: Overview of the 5th Author Profiling
Task at PAN 2017: Gender and Language Variety Identification in Twitter. In: Cappellato,
L., Ferro, N., Goeuriot, L., Mandl, T. (eds.) Working Notes Papers of the CLEF 2017
Evaluation Labs. CEUR Workshop Proceedings, vol. 1866. CLEF and CEUR-WS.org (Sep
2017), http://ceur-ws.org/Vol-1866/
23. Rangel Pardo, F., Rosso, P., Verhoeven, B., Daelemans, W., Potthast, M., Stein, B.:
Overview of the 4th Author Profiling Task at PAN 2016: Cross-Genre Evaluations. In:
Working Notes Papers of the CLEF 2016 Evaluation Labs. CEUR Workshop Proceedings,
vol. 1609. CLEF and CEUR-WS.org (Sep 2016), http://ceur-ws.org/Vol-1609/
24. Ronan, C., Jason, W.: A unified architecture for natural language processing: Deep neural
networks with multitask learning. In: Proceedings of the 25th International Conference on
Machine Learning. pp. 160–167. ICML ’08, ACM, New York, NY, USA (2008),
http://doi.acm.org/10.1145/1390156.1390177
25. Rong-En, F., Kai-Wei, C., Cho-Jui, H., Xiang-Rui, W., Chih-Jen, L.: Liblinear: A library for
large linear classification. J. Mach. Learn. Res. 9, 1871–1874 (Jun 2008),
http://dl.acm.org/citation.cfm?id=1390681.1442794
26. Stamatatos, E., Rangel, F., Tschuggnall, M., Kestemont, M., Rosso, P., Stein, B., Potthast,
M.: Overview of PAN-2018: Author Identification, Author Profiling, and Author
Obfuscation. In: Bellot, P., Trabelsi, C., Mothe, J., Murtagh, F., Nie, J., Soulier, L., Sanjuan,
E., Cappellato, L., Ferro, N. (eds.) Experimental IR Meets Multilinguality, Multimodality,
and Interaction. 9th International Conference of the CLEF Initiative (CLEF 18). Springer,
Berlin Heidelberg New York (Sep 2018)
27. Tellez, E., Miranda-Jiménez, S., Graff, M., Moctezuma, D.: Gender and language-variety
Identification with MicroTC—Notebook for PAN at CLEF 2017. In: Cappellato et al. [22],
http://ceur-ws.org/Vol-1866/
28. Thorsten, J.: Text categorization with support vector machines: Learning with many
relevant features. In: Claire, N., Céline, R. (eds.) Machine Learning: ECML-98. pp.
137–142. Springer Berlin Heidelberg, Berlin, Heidelberg (1998)
29. Zhang, X., Zhao, J.J., LeCun, Y.: Character-level convolutional networks for text
classification. CoRR abs/1509.01626 (2015),
http://arxiv.org/abs/1509.01626