=Paper=
{{Paper
|id=Vol-2125/paper_74
|storemode=property
|title=Author Profiling: Gender Prediction from Tweets and Images: Notebook for PAN at CLEF 2018
|pdfUrl=https://ceur-ws.org/Vol-2125/paper_74.pdf
|volume=Vol-2125
|authors=Yaakov HaCohen-Kerner,Yair Yigal,Elyashiv Shayovitz,Daniel Miller,Toby Breckon
|dblpUrl=https://dblp.org/rec/conf/clef/HaCohen-KernerY18
}}
==Author Profiling: Gender Prediction from Tweets and Images: Notebook for PAN at CLEF 2018==
https://ceur-ws.org/Vol-2125/paper_74.pdf
Author Profiling: Gender Prediction
from Tweets and Images
Notebook for PAN at CLEF 2018
Yaakov HaCohen-Kerner1, Yair Yigal1, Elyashiv Shayovitz1, Daniel Miller1, and
Toby Breckon2
1
Department of Computer Science, Jerusalem College of Technology, Lev Academic Center
21 Havaad Haleumi St., P.O.B. 16031, 9116001 Jerusalem, Israel
2
Department of Computer Science, the Palatine Centre, Durham University, Stockton Road, Durha, DH1 3LE, UK
kerner@jct.ac.il, yigalyairn@gmail.com, elyashiv12@gmail.com,
3danmi@gmail.com, toby.breckon@durham.ac.uk
Abstract. Author profiling deals with identification of various details about the
author of the text (e.g., age, cultural background, gender, native language,
personality). In this paper, we describe the participation of our teams (yigal18
and miller18, both teams contain the same people, but in another order) in the
PAN 2018 shared task on author profiling, identifying authors’ gender where for
each author, 100 tweets and 10 images are provided. The authors were grouped
by the language of their tweets: English, Spanish, and Arabic. In this paper, we
describe our pre-processing, feature sets, machine learning methods and accuracy
results. The best results using the textual features were achieved using the MLP
method after applying the L normalization and using 9,000 word unigrams for
English, 10,000 word unigrams and one orthographic feature for Spanish, and
7,000 word unigrams and one orthographic feature for Arabic. We also tried
various additional feature sets, including style-based feature sets. In most of the
cases, these features did not improve the results and in a few cases even hurt the
results. The best result (61.54%) for the visual features was obtained by the LR
method using all the features (SIFT & Color & VGG) and the best basic feature
set is the VGG. The best result for the combined features was achieved using
model2 (miller18) with 0.75 as a weight to the best textual model and a weight
of 0.25 for NN Classifier (Keras) using only the 1000 VGG features.
Keywords: Author Profiling, Gender Classification, Content-based Features, Style-
based Features, Images, Supervised Machine Learning, Tweets, Visual Features.
1 Introduction
Author profiling deals with analysis of a given text while inferring various information
about the author of the text (e.g., age, cultural background, gender, native language,
personality). This problem is of growing importance all over the world. Important and
interesting applications can be found in business intelligence, forensics, psychology,
and security. A linguistic analysis of a given text can help to identify certain
characteristics of the author. For example, companies would like to know, based on the
analysis of online product reviews, the demographics of people who like or dislike their
products.
In this paper, we describe the participation of our teams (yigal18 and miller18, both
teams contain the same people, but in another order) in the PAN 2018 shared task on
author profiling. More specifically, the shared task is on gender identification of authors
from their tweets and images (100 tweets and 10 images are provided for each author).
The addressed languages are English, Spanish and Arabic.
We consider application of several supervised machine learning (ML) methods and
various types of features for the gender identification. Content-based features and style-
based features are extracted from the tweets and visual features are extracted from the
images.
The rest of the paper is organized as follows. Section 2 provides background and
presents some related work on text classification in general and author profiling in
particular. Section 3 introduces the feature sets (content-based, style-based, and visual)
that we have implemented and used in our extensive experiments. Section 4 presents
the experimental setup, the experimental results using three modalities (textual feature
and / or visual features) for corpora written in three languages and their analysis.
Finally, Section 5 summarizes and suggests ideas for future research.
2 Related Work
2.1 Text classification
TC is the supervised learning task of assigning natural language text documents to one
or more predefined categories [31]. There are two main types of TC: topic-based
classification and style-based classification. An example of a topic-based classification
application is classifying news articles to various categories such as Business-Finance,
Lifestyle-Leisure, Science-Technology and Sports, that were downloaded from three
well-known news web-sites (BBC, Reuters, and TheGuardian) [28].
An example of a style-based classification application is classification based on
different literary genres, e.g., action, comedy, crime, fantasy, historical, political, saga,
and science fiction [25, 12].
These two classification types often require different types of feature sets to achieve
the best performance. Topic-based classification is typically performed using word uni-
grams and/or n-grams (n > 2) ([2, 15]. Style-based classification is typically performed
using linguistic features such as quantitative features, orthographic features, part of
speech (POS) tags, function words, and vocabulary richness features [25, 16, 17,12].
2.2 Author profiling
Author profiling can be viewed as a sub-task of TC. The author profiling task is of
growing importance during recent years. Various author profiling applications are
found in business intelligence, forensics, psychology, and security systems. The general
aim of an author profiling task is to determine various |demographic information about
the text’s author(s), e.g., age, cultural background, gender, native language and/or
dialect, and various personality traits. In this paper, we will limit ourselves to gender
studies because the PAN 2018 shared task is on gender identification.
The gender classification might be viewed as one of the simplest classification tasks.
The classification decision can be binary and a relatively large amount of data can be
collected. However, such a classification system can be effective only if the writing
style between genders does differ [10] and if such stylistic differences can be detected
[23].
In contrast to other demographic traits, the link between gender and word use has
been extensively studied [34]. Differences in women's and men's language have
received relatively high attention within the scientific community as well as in the
popular media. However, early studies on gender classification, mainly on formal texts
and blogs, reported on accuracies around 75%-80% in most cases [26, 1, 44, 19, 7].
Lakoff [27] found that women use less assertive speech that manifests itself in a
higher degree of politeness, less swearing, more frequent tag questions (e.g., "it is …,
isn't it?"), more intensifiers (e.g., really, so), and more hedges (e.g., sort of, perhaps,
maybe; also known as qualifiers or uncertainty words).
Mulac et al. [32] summarized the findings of more than 30 empirical studies and
reported that typical female language features contain intensive adverbs (e.g., really,
so), references to emotions, uncertainty verbs (seems to, maybe), negations (e.g., not,
never), and hedges; while typical male language features include references to quantity,
judgmental adjectives (e.g., good, dumb), elliptical sentences ("Great picture."),
directives ("Write that down."), and "I" references.
Schler et al. [45] found that men’s writing is more related to money, job and TV,
while women’s writing is more related to family, sex and eating. Argamon et al. [3]’s
experimental results regarding gender showed that the style-based features that proved
to be most useful for gender discrimination are determiners and prepositions (markers
of male writing) and pronouns (markers of female writing). The content features that
prove to be most useful for gender discrimination are words related to technology
(male) and words related to personal life or relationships (female).
Pennebaker [35] found that women tend to use more personal pronouns and words
referring to emotions. By the contrary, men tend to use more articles, long words
(defined as words with more than 6 characters), nouns, and prepositions.
Rangel et al. [39] presented in their overview paper the framework and the results
for the Author Profiling task at PAN 2015, which dealt with the identification of age,
gender, and personality traits of Twitter users. In comparison to previous years of PAN
[37, 38] the PAN-15 systems achieved significantly higher accuracy values for gender
identification. This may suggest that, irrespective the shorter length of individual tweets
and their informality, the number of tweets per author is sufficient to profile age and
gender with high accuracy. Regarding the features, it was not clear which ones were
the most important ones, because the high number of different ones used and combined
by the teams. Some of the best teams used the Second Order Representation and other
were based on n-gram representation.
A similar phenomenon occurred in the gender classification tasks in [41]. It was
difficult to highlight the contribution of any particular feature since the teams used
many of them. The second order representation was used by teams that achieved first
positions in some of the tasks. Likewise, the distributed representations achieved the
first position in gender identification on the Dutch final evaluation.
The best resulting approaches that took part in the gender classification tasks in
PAN 2017 [42] took advantage from combinations of n-grams, other content-based
features, and style-based features. The best final average gender ranking (for English,
Portuguese, and Spanish) shows that the best overall result (82.53%) has been obtained
by Basile et al. [4], who used the scikit-learn LinearSVM implementation trained with
combinations of character 3- to 5-grams and word 1- to 2-grams with TF-IDF weighting
with sublinear term frequency scaling.
3 Features
In this section, we present the various types of features that we applied for the gender
identification. Content-based features and style-based features are extracted from the
tweets and the visual features are extracted from the images.
3.1 Textual feature sets
In this sub-section, we describe the basic super-sets and sets of the textual features we
use for our classification experiments: content-based features and style-based features.
Our content-based features include various n-grams sets, where each one of them is
defined by the following template: number_k-n_type where number is the number of
the features in the set, k is the size of the wanted skip (0 – no skip, 1 – skip of one unit,
2 – skip of 2 units, …), n is code of the grams (1 for unigrams 2 for bigrams, 3 for
trigrams, …), and type is W for words or C for characters. All values are represented
by TF-IDF values. The specific various n-grams sets that were applied will be presented
later in the framework of the experiments.
Our style-based features include the following feature sets: Quantitative features,
Orthographic features, Gender features that contain letters or words based on the special
tweet text features that appear in Figure 7 in Burger et al. [7], and Gender features that
do not contain any letter. The Quantitative set includes 3 features: number of characters
in the tweet, number of words in the tweet, and average number of characters in a word.
The Orthographic set includes only one feature - the total frequency of the following
characters )"/ \ ) - : # @ $ & . , ? !"( normalized by the number of characters in the
tweet. The Gender features that do not contain any letter include the frequency of each
following character strings: ,): ,_! ,:_ ,!y!, !_i, _i, ooo, !_i, <3, :)$, :(, _:)$, !$, _<3,
_b, y:_) normalized by the number of characters in the tweet.
3.2 Visual feature sets
In this sub-section, we describe the three basic sets of visual features we used for our
classification experiments: SIFT1, Color2, and VGG3.
The SIFT (Scale Invariant Feature Transform) algorithm was developed by Lowe
[29, 30]. The SIFT algorithm detects and describes local features in images. It consists
of two sub-processes. The first sub-process is a process that detects interest points in
an image. Interest points are where the signal in 2D space has variation that exceeds
some threshold criterion and is superior to simple edge detection. The second sub-
process creates a vector like descriptor. To create scale invariance, the interest points
are scanned at a wide range of scales. For our experiments, we extract from each image
at least 500 key-points, and describe each interest point by 128 features. In addition, we
create by K-means algorithm a bag of 0111 visual words. For each image, we create a
histogram of those words.
The Color algorithm extracts the colors in an image (by the RGB model) and for
each color, indicates its frequency (pixels wise) in the image. Using the Colorgram4
library we extract the 500 most frequent colors in the images, and for each image we
build a histogram of the colors in it.
VGG5 (Oxford Visual Geometry Group) is a group of researchers who built a model
containing deep convolutional networks for large-scale image recognition. Using the
ORB6 (Oriented FAST and Rotated BRIEF) system, we extract from each image at least
500 key-points. Then, we describe each key-point by a vector of VGG features. For
each user we have ten images; therefore, we have at least 5000 VGG vectors for each
user. Then we create for each user a histogram containing 1000 visual words .
4 Experimental Setup and Results
The PAN CLEF 2018 [47] launched an evaluation campaign. Twenty-three teams have
participated in this campaign. Each team has proposed its algorithm, which has been
evaluated using the TIRA platform [36]. The algorithms and the results of the
participated teams have been overviewed in Rangel et al. [43].
General approach: Our approach to authorship profiling is to apply supervised ML
methods to TC as was suggested by Sebastiani [44]. The process is as follows. First,
given a corpus of training documents, where each document is labeled as either ‘male’
1
https://github.com/abidrahmank/OpenCV2-Python
Tutorials/blob/master/source/py_tutorials/py_feature2d/py_sift_intro/py_sift_intro.rst
2
https://github.com/obskyr/colorgram.py/blob/master/colorgram/colorgram.py
3
https://docs.opencv.org/trunk/d6/d00/classcv_1_1xfeatures2d_1_1VGG.html
4
https://pypi.org/project/colorgram.py/#description
5
https://arxiv.org/abs/1409.1556
6
http://opencv-python-tutroals.readthedocs.io/en/latest/py_tutorials/py_feature2d/py_orb/py_orb.html
or ‘female’, we processed each document to produce values for various combinations
of features from different types of features sets: content-based features, style-based
features, and visual features. Second, we applied several popular ML methods on the
generated combinations of features. Third, we tried additional combinations of features
and parameter tuning. Finally, the best model(s) for the training data were tested on
out-of-training data (i.e., test data).
Preprocessing: There is a widespread variety of text preprocessing types such as:
conversion of uppercase letters into lowercase letters, html object removal, stopword
removal, punctuation mark removal, reduction of different sets of emoticon labels to a
reduced set of wildcard characters, replacement of HTTP links to wildcard characters,
word stemming, word lemmatization, correction of common misspelled words, and
reduction of replicated characters. Not all of them are considered as effective by all TC
researchers. Many systems use only a small number of simple preprocessing types (e.g.,
conversion of all the uppercase letters into lowercase letters and / or stopword removal).
In our classification experiments, we tried the following text preprocessing types: L
– converting uppercase letters into lowercase letters, U – URL link removal, P –
punctuation mark removal, S – stopword removal, C – Error Correction, and T –
stemming. The application of the S preprocessing type deletes all instances of 423
stopwords for English text (421 stopwords from Fox [11] plus the letters “x” and “z”
that are not found in Fox [11], yet are included in many other stopword lists).
ML methods: We applied four ML methods: MLP– Multilayer Perceptron7,
LinearSVC – SVM with a linear kernel8, LR - Logistic regression9, and RF - Random
Forest10.
A brief description of these ML methods follows: MLP is a feedforward neural
network ML method [21] where artificial neural network (ANN) can be viewed as a
weighted directed graph in which nodes are artificial neurons and directed edges (with
weights) are connections from the outputs of neurons to the inputs of neurons. Support
vector machine (SVM, also called support vector network) [9] is a model that assigns
examples to one of two categories, making it a non-probabilistic binary linear classifier.
LinearSVC is SVM with a linear kernel, which is recommended for TC because most
of TC problems are linearly separable [22] and training a SVM with a linear kernel is
faster compared to other kernels. Logistic regression (LR) is a variant of a statistical
model that tries to predict the outcome of a categorical dependent variable (i.e., a class
label) [8, 20]. Random Forest (RF) is an ensemble learning method for classification
and regression [6]. RF operates by constructing a multitude of decision trees at training
time and outputting classification for the case at hand. RF combines the “bagging”
idea presented by Breiman [5] and random selection of features introduced by Ho [18]
to construct a forest of decision trees.
Tools and information sources: We used the following tools:
Scikit-learn11 - a library for ML methods
7
http://scikit-learn.org/stable/modules/neural_networks_supervised.html
8
http://scikit-learn.org/stable/modules/generated/sklearn.svm.LinearSVC.html
9
http://scikit-learn.org/stable/modules/generated/sklearn.linear_model.LogisticRegression.html
10
http://scikit-learn.org/stable/modules/generated/sklearn.ensemble.RandomForestClassifier.html
11
http://scikit-learn.org/stable/index.html
NLTK12- a library that produces the various n-gram features
Numpy13 - a library that performs fast mathematical processing
Autocorrect14- a library that automatically corrects spelling errors.
4.1 Experimental results using the textual feature sets
The first TC experiments were performed for the English corpus using the four ML
methods without parameter tuning and any normalization types using only sets of word
unigrams. The results are shown in Table 1. The best result is bolded.
Table 1. English corpus - Accuracy results using sets of word unigrams.
Features MLP LinearSVC LR RF
1000 Word Unigrams 73.06061 75.55556 73.13131 65.77778
2000 Word Unigrams 73.93939 75.65657 73.53535 62.0202
3000 Word Unigrams 74.94949 75.65657 73.53535 64.34343
4000 Word Unigrams 75.85859 76.56566 74.34343 65.15152
5000 Word Unigrams 76.26263 75.85859 74.0404 63.53535
6000 Word Unigrams 76.56566 75.85859 73.43434 66.56566
7000 Word Unigrams 76.56566 75.45455 73.83838 65.45455
8000 Word Unigrams 75.85859 76.26263 73.73737 63.53535
9000 Word Unigrams 78.28283 76.76768 73.53535 63.53535
10000 Word Unigrams 77.17172 76.36364 73.63636 65.55556
20000 Word Unigrams 76.46465 76.06061 73.73737 62.72727
30000 Word Unigrams 76.46465 75.9596 73.73737 65.65657
40000 Word Unigrams 75.55556 75.85859 73.53535 62.12121
50000 Word Unigrams 76.26263 76.16162 73.73737 66.56566
Similar experiments have been performed on various sets of word bigrams and the
best accuracy result 74.95 was obtained using 20,000 word bigrams.
In our classification experiments, we tried various combinations of text
preprocessing types: L – converting uppercase letters into lowercase letters, U – URL
link removal, P – punctuation mark removal, S – stopword removal, C – Error
Correction, and T – stemming. Only the lowercase preprocessing improved the TC
results. Therefore, in our final experiments, we applied only this preprocessing type.
We have also tried various sets of skip character/word n-grams and various style-
based feature sets (Quantitative features, Orthographic features, Gender features that
contain letters or words, and Gender features that do not contain any letter) in our
feature combinations. However, they did not improve the accuracy results.
12
https://www.nltk.org/
13
http://www.numpy.org/
14
https://github.com/phatpiglet/autocorrect
Furthermore, we also applied Principal component analysis (PCA) [24]. However, it
did not improve the accuracy results.
Similar results have been achieved for the corpora in Spanish and Arabic.
Eventually, based on the results of the previous experiments, we chose to apply the
MLP method with the following features for the three languages:
English: combination of 9,000 word unigrams with the L normalization
Spanish: combination of 10,000 word unigrams and one orthographic feature with
the L normalization for
Arabic: combination of 7,000 word unigrams and one orthographic feature with
the L normalization
This setting was used both by yigal18 and miller18. That is to say, yigal18 and
miller18 used the same setting for the experiments on the test data for the textual
features.
4.2 Experimental results using the visual feature sets
In the classification experiments using the visual features under the framework of
scikit-learn15 [33], we apply seven versions of five ML methods: (1) Decision tree:
default, using mean scaling16; (2) Logistic regression (C=1e-4, max_iter=1e5, using
mean scaling); (3) SVM ('poly' kernel): C=1e-3, gamma=.01, using mean scaling; (4)
SVM ('rbf' kernel): C=1e-4, gamma=.01; (5) SDGClassifier: epsilon=0.7,
max_iter=10000;(6) NN Classifier (version 1): activation='relu', alpha=1e2,
hideen_layers=(64, ), epsilon=0.8, using mean scaling; and (7) NN Classifier (version
2): activation='relu', epoches=30, hideen_layers=(16,1, ).
We used the following sets of visual features: SIFT, SIFT & Color, VGG, VGG &
SIFT& Colors, VGG & Colors, VGG & SIFT. Table 2 presents the final results of all
possible feature sets combinations (including the basic features sets alone) using the
seven versions of five ML methods on the biggest set (2000 training samples, 1000
testing samples). The three best results are in bold.
15
http://scikit-learn.org/stable/index.html
16
http://scikit-learn.org/stable/modules/generated/sklearn.preprocessing.StandardScaler.html
Table 2. The final results of all combinations of visual feature sets using the ML methods.
Feat. Comb. SIFT Color VGG SIFT & SIFT & Color SIFT &
Color VGG & Color &
ML method VGG VGG
Decision Tree 48.58 50.90 50.79 50.60 51.31 52.81 53.23
Logistic 58.99 54.14 61.06 58.48 60.99 58.52 61.54
Regression
SGCClassifier 54.75 53.73 56.21 52.12 56.27 55.10 58.00
SVM 60.2 51.33 59.65 58.99 60.16 57.63 60.14
('poly' kernel)
SVM 48.99 56.50 46.25 48.99 49.03 48.59 47.22
('rbf' kernel)
NN Classifier 57.37 53.23 58.88 58.28 58.18 57.11 60.29
(Sklearn)
NN Classifier 56.77 53.83 61.13 57.78 51.12 56.59 59.69
(Keras)
The best accuracy result in Table 2 is 61.54%, which was obtained by the LR method
using all the features (SIFT & Color & VGG). LR obtained the first place for three
feature combinations (columns) and SVM ('poly' kernel) obtained the first place for two
other feature combinations. The best basic feature set is the VGG (best result in 4 out
7 ML methods). Two ML methods (LR and NN Classifier (Keras)) obtained results
above 61% using the VGG features that are very close to the best result in the table that
was obtained by all the features. This means that the VGG probably contributes the
most and the two other feature sets almost did not contribute.
For the first model (yigal18) we used the following features: 1000 VGG features,
1000 SIFT features, and 500 Color features. The learning method was a Logistic
Regression (C=1e-2), with mean scaling.
For the second model (miller18), we used only the 1000 VGG features. The chosen
ML method was the NN Classifier (Keras) (with 64 hidden units at one layer).
4.3 Experimental results using both the textual and visual feature sets
For this experiment, we used two different models (yigal18 and miller18). The
construction of each model was performed according to the same general method: the
data was separately classified by the visual and textual features, and using the next
formula we combined the two classifications into a final classification.
𝐶𝑜𝑚𝑏𝑎𝑖𝑛𝑒𝑑𝑃𝑟𝑜𝑏(𝑐𝑙𝑎𝑠𝑠𝑖 ) = 𝛼 ∗ 𝑃𝑡𝑒𝑥𝑡𝑢𝑎𝑙 (𝑐𝑙𝑎𝑠𝑠𝑖 ) + (1 − 𝛼) ∗ 𝑃𝑣𝑖𝑠𝑢𝑎𝑙 (𝑐𝑙𝑎𝑠𝑠𝑖 )
For each model, we used the same setting for the textual features (see sub-section
4.1) and the MLP method. Different settings of the visual features and different ML
methods have been applied by the two models.
The first model (yigal18) applied Logistic Regression on the following visual
features: 1000 VGG features, 1000 SIFT features, and 500 Color features. The value of
α was set to 0.5, which means that we gave equal weighs for the textual and visual
components.
The second model (miller18) used as its visual features only the 1000 VGG features.
The chosen ML method was the NN Classifier (Keras) (with 64 hidden units at one
layer). The textual features remain the same since their results in the first model were
relatively high. The value of α was set to 0.75 (i.e., giving higher weight for textual
features because they obtained better results).
The results of the two models yigal18 and miller18 are presented in Tables 3 and 4,
respectively. The best results for the other modalities are in bold.
Table 3. The accuracy results (in %) of the first model (yigal18).
Textual Visual Combined
English 79.11 49.42 78.89
Arabic 75.9 51 75.7
Spanish 76.5 50.27 75.91
Table 4. The accuracy results (in %) of the second model (miller18) using the MLP method.
Textual Visual Combined
English 79.11 51.74 79.47
Arabic 75.9 49.7 75.7
Spanish 76.5 49.23 76.23
The comparison of the yigal18 (model1, Table 3) and miller18 (model2, Table 4)
leads to the following findings and perhaps to the following conclusions. (1) Regarding
the visual results, there is a decrease in two languages (Spanish & Arabic) and an
increase in English. A possible explanation is that using all the visual features (SIFT &
Color& VGG) leads to slightly better results than using only the VGG features. (2)
Regarding the combined results, there is an increase in two languages (English &
Spanish) and no change for Arabic. The improvements in miller18 (model2) are
probably because of the upgrade of alpha from 0.5 to 0.75. That is to say, i.e., higher
weight for the textual features leads to better combined results. (3) The second visual
model run faster due to the lower number of features (1000 VGG features), in contrast
to the first model (all the 2500 visual features). (4) The result of the combined features
for model2 is better than the result of the textual features for model2. This finding
means that the visual features slightly contribute to improve the result obtained by the
textual features. This phenomenon did not occur in all other cases in both models.
5 Summary and Future Work
In this paper, we describe our participation in the PAN 2018 shared task on author
profiling, identifying authors’ gender, based on their images and tweets. We tried
various pre-processing types, a widespread variety of feature sets, and four ML
methods.
The best results using the textual features were achieved using the MLP method after
applying the L normalization and using 9,000 word unigrams for English, 10,000 word
unigrams and one orthographic feature for Spanish, and 7,000 word unigrams and one
orthographic feature for Arabic.
The best result (61.54%) for the visual features was obtained by the LR method using
all the features (SIFT & Color & VGG). The best basic feature set is the VGG. Two
other results (also above 61%) were obtained while applying LR and NN (Keras) using
the VGG features. This means that the VGG probably contributes the most and the two
other feature sets almost did not contribute.
The best results using the combined features were achieved using model2 with 0.75
as a weight to the textual features and 0.25 as a weight to the visual features. For the
textual features we used the best model that was used also in model1 (sub-section 5.3).
For the visual features of the second model, we use only the 1000 VGG features. The
chosen ML method was the NN Classifier (Keras) (with 64 hidden units at one layer).
Future research proposals include: (1) applying additional combinations of feature
sets; (2) tuning each model (textual and visual) separately; (3) tuning the value of alpha
to find the best combined model; (4) developing a new combined model performing
multi-modal fusion using the textual and visual modalities; (5) applying various deep
neural models; and (6) building model that will perform author profiling in general and
gender prediction in particular using keyphrases [13,14] that distinguish each of classes
in general and each gender in paticular.
Acknowledgments. This work was partially funded by the Jerusalem College of
Technology (Lev Academic Center) and we gratefully acknowledge its support.
References
1. Argamon, S., Koppel, M., Fine, J., Shimoni, A. R.: Gender, genre, and writing style in
formal written texts. TEXT-THE HAGUE THEN AMSTERDAM THEN BERLIN-
, 23(3), 321-346 (2003).
2. Argamon, S., Whitelaw, C., Chase, P., Hota, S. R., Garg, N., Levitan, S.: Stylistic text
classification using functional lexical features: Research articles. Journal of the
American Society for Information Science and Technology. 58, 6, 802–822 (2007).
3. Argamon, S., Koppel, M., Pennebaker, J. W., Schler, J.: Automatically profiling the
author of an anonymous text. Communications of the ACM, 52(2), 119-123 (2009).
4. Basile, A., Dwyer, G., Medvedeva, M., Rawee, J., Haagsma, H., Nissim, M.: N-GrAM:
New Groningen Author-profiling Model. arXiv preprint arXiv:1707.03764 (2017).
5. Breiman, L.: Bagging predictors. Univ. California Technical Report No. 421. (1994).
6. Breiman, L.: Random forests. Machine learning, 45(1), 5-32 (2001).
7. Burger, J. D., Henderson, J., Kim, G., Zarrella, G.: Discriminating gender on Twitter.
In Proceedings of the conference on empirical methods in natural language processing
(pp. 1301-1309). Association for Computational Linguistics (2011).
8. Cessie, S. Le, Van Houwelingen, J. C.: Ridge estimators in logistic regression, Applied
statistics, pp. 191-201 (1992).
9. Cortes, C., Vapnik, V.: Support-vector networks. Machine learning, 20(3), 273-297
(1995).
10. Eckert, P., McConnell-Ginet, S.: Language and gender. Cambridge University Press
(2013).
11. Fox, C.: A stop list for general text. In Acm sigir forum (Vol. 24, No. 1-2, pp. 19-21).
ACM (1989).
12. Gianfortoni, P., Adamson, D., Rosé, C. P.: Modeling of stylistic variation in social
media with stretchy patterns. In Proceedings of the First Workshop on Algorithms and
Resources for Modelling of Dialects and Language Varieties (pp. 49-59). Association
for Computational Linguistics (2011).
13. HaCohen-Kerner, Y., Gross, Z., Masa, A.: Automatic extraction and learning of
keyphrases from scientific articles. In Int. Conf. on Intelligent Text Processing and
Computational Linguistics, Springer, Berlin, Heidelberg, pp. 657-669 (2005).
14. HaCohen-Kerner, Y., Stern, I., Korkus, D., Fredj, E.: Automatic machine learning
of keyphrase extraction from short html documents written in
Hebrew. Cybernetics and Systems: An International Journal, 38(1), 1-21 (2007).
15. HaCohen-Kerner, Y., Mughaz, D., Beck, H., Yehudai, E.: Words as classifiers of
documents according to their historical period and the ethnic origin of their authors.
Cybernetics and Systems: An International Journal, 39(3), 213-228 (2008).
16. HaCohen-Kerner, Y., Beck, H., Yehudai, E., Rosenstein, M., Mughaz, D.: Cuisine:
Classification using stylistic feature sets &/or name-based feature sets. Journal of the
American Society for Information Science and Technology 61 (8), 1644–57 (2010A).
17. HaCohen-Kerner, Y., Beck, H., Yehudai, E., Mughaz, D.: Stylistic feature sets as
classifiers of documents according to their historical period and ethnic origin. Applied
Artificial Intelligence 24 (9), 847–62 (2010B).
18. Ho, T. K.: Random Decision Forests. Proceedings of the 3rd Int. Conference on
Document Analysis and Recognition, Montreal, QC, 14–16 August 1995. 278–282
(1995).
19. Holmes, J., Meyerhoff, M. (Eds.).: The handbook of language and gender (Vol. 25).
John Wiley & Sons (2008).
20. Hosmer Jr, D. W., Lemeshow, S., Sturdivant, R. X.: Applied logistic regression (Vol.
398). John Wiley & Sons (2013).
21. Jain, A. K., Mao, J., Mohiuddin, K. M.: Artificial neural networks: A
tutorial. Computer, 29(3), 31-44 (1996).
22. Joachims, T.: Text categorization with support vector machines: Learning with many
relevant features. In European conference on machine learning (pp. 137-142).
Springer, Berlin, Heidelberg (1998).
23. Jockers, M. L., Witten, D. M.: A comparative study of machine learning methods for
authorship attribution. Literary and Linguistic Computing, 25(2), 215-223 (2010).
24. Jolliffe, I. T.: Principal component analysis and factor analysis. In Principal component
analysis (pp. 115-128). Springer, New York, NY (1986).
25. Kessler, Brett, Nunberg, Geoffrey, Hinrich Schutze.: Automatic detection of text
genre. In P. R. Cohen & W. Wahlster (Eds.), In Proc. of the 35th annual meeting of the
ACL and 8th conf. of the European chap. of the Assoc. for Computational Linguistics.
32-38, Somerset, New Jersey: Association for Computational Linguistics (1997).
26. Koppel, M., Argamon, S., Shimoni, A. R.: Automatically categorizing written texts by
author gender. Literary and linguistic computing, 17(4), 401-412 (2002).
27. Lakoff, R.: Language and woman's place. Language in society, 2(1), 45-79 (1973).
28. Liparas, D., HaCohen-Kerner, Y., Moumtzidou, A., Vrochidis, S., Kompatsiaris, I.
News articles classification using Random Forests and weighted multimodal features.
In Information Retrieval Facility Conference, Springer, Cham, pp. 63-75, (2014).
29. Lowe, David G.: Object recognition from local scale-invariant features. In Proc. of the
Int. Conf. on Computer Vision. 2., pp. 1150–1157 (1999).
30. Lowe, David G.: Distinctive image features from scale-invariant
keypoints. International journal of computer vision, 60(2), 91-110 (2004).
31. Meretakis, D., Wuthrich, B.: Extending naive Bayes classifiers using long itemsets,
Proc. of the 5th ACM-SIGKDD Int. Conf. Knowledge Discovery, Data Mining
(KDD'99), San Diego, USA, 165-174 (1999).
32. Mulac, A., Bradac, J. J., Gibbons, P.: Empirical support for the gender-as-culture
hypothesis: An intercultural analysis of male/female language differences. Human
Communication Research, 27(1), 121-152 (2001).
33. Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., ...
Vanderplas, J.: Scikit-learn: Machine learning in Python. Journal of machine learning
research, 12(Oct), 2825-2830 (2011).
34. Pennebaker, J. W., Mehl, M. R., Niederhoffer, K. G.: Psychological aspects of natural
language use: Our words, our selves. Annual review of psychology, 54(1), 547-577
(2003).
35. Pennebaker, J.W.: The secret life of pronouns: What our words say about us. New
York: Bloomsbury Press (2011).
36. Potthast, M., Gollub, T., Rangel, F., Rosso, P., Stamatatos, E., Stein, B.: Improving the
Reproducibility of PAN's Shared Tasks: Plagiarism Detection, Author Identification,
and Author Profiling. In: Kanoulas, E., Lupu, M., Clough, P., Sanderson, M., Hall, M.,
Hanbury, A., Toms, E. (eds.) Information Access Evaluation meets Multilinguality,
Multimodality, and Visualization. 5th International Conference of the CLEF Initiative
(CLEF 14). pp. 268–299. Springer, Berlin Heidelberg New York (Sep 2014).
37. Rangel, F., Rosso, P., Koppel, M., Stamatatos, E., Inches, G.: Overview of the author
profiling task at PAN 2013. In CLEF Conference on Multilingual and Multimodal
Information Access Evaluation, CELCT, pp. 352-365 (2013).
38. Rangel, F., Rosso, P., Chugur, I., Potthast, M., Trenkmann, M., Stein, B., ...
Daelemans, W.: Overview of the 2nd author profiling task at pan 2014. In CLEF 2014
Evaluation Labs and Workshop Working Notes Papers, Sheffield, UK, 2014, pp. 1-30
(2014).
39. Rangel, F., Rosso, P., Potthast, M., Stein, B., Daelemans, W.: Overview of the 3rd
Author Profiling Task at PAN 2015. In CLEF p. 2015 (2015).
40. Rangel, F., Franco-Salvador, M., Rosso, P.: A low dimensionality representation for
language variety identification. In International Conference on Intelligent Text
Processing and Computational Linguistics Springer, Cham, pp. 156-169 (2016A).
41. Rangel, 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/Balog, Krisztian [edit.]; et al., pp. 750-784 (2016).
42. Rangel, F., Rosso, P., Potthast, M., Stein, B.: Overview of the 5th author profiling task
at pan 2017: Gender and language variety identification in twitter. Working Notes
Papers of the CLEF (2017).
43. 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)
44. Sebastiani, F.: Machine learning in automated text categorization. ACM
computing surveys (CSUR), 34(1), 1-47 (2002).
45. Schler, J., Koppel, M., Argamon, S., Pennebaker, J. W.: Effects of age and gender on
blogging. In AAAI spring symposium: Computational approaches to analyzing
weblogs, Vol. 6, pp. 199-205 (2006).
46. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image
recognition. arXiv preprint arXiv:1409.1556 (2014).
47. 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 Int. Conference of the CLEF Initiative (CLEF 18).
Springer, Berlin Heidelberg New York (Sep 2018).