=Paper=
{{Paper
|id=Vol-3159/T7-2
|storemode=property
|title=Ensembled Feature Selection for Urdu Fake News Detection
|pdfUrl=https://ceur-ws.org/Vol-3159/T7-2.pdf
|volume=Vol-3159
|authors=Fazlourrahman Balouchzahi,Hosahalli Lakshmaiah Shashirekha,Grigori Sidorov
|dblpUrl=https://dblp.org/rec/conf/fire/BalouchzahiSS21b
}}
==Ensembled Feature Selection for Urdu Fake News Detection==
https://ceur-ws.org/Vol-3159/T7-2.pdf
Ensembled Feature Selection for Urdu Fake News
Detection
Fazlourrahman Balouchzahi1 , Hosahalli Lakshmaiah Shashirekha2 and
Grigori Sidorov1
1 Instituto Politécnico Nacional, Centro de Investigación en Computación, CDMX, Mexico
2 Department of Computer Science, Mangalore University, Mangalore, India
Abstract
Identifying fake news shared on social media is a vital task due to its immense effects in a negative
way on the society, community, an individual or whoever is the target. Controlling and managing the
fake news shared on social media manually is an impractical task due to the increasing number of
social media users, increasing volume of fake news and the speed in which the fake news spreads on
social media. Hence, there is a great demand for the automatic identification of fake news quickly and
efficiently. Most of the fake news detection works carried out focus on resource rich languages like
English and Spanish leaving the under-resourced languages like Urdu and many Indian languages less
attended or unattended. UrduFake 2021 - a shared task in Forum for Information Retrieval Evaluation
(FIRE) 2021 promotes detecting fake news in Urdu - an under-resourced language. This paper presents
the description of the model proposed and submitted by our team MUCIC to UrduFake 2021 which aims
to classify Urdu news article into one of the two categories, namely: Fake and Real. The major focus of
this work is on feature engineering part to enhance the performance of traditional Machine Learning
(ML) classifiers using very simple features such as word and char n-grams. Three Feature Selection (FS)
algorithms, namely: Chi-square, Mutual Information Gain (MIG), and f_classif are ensembled to select
the top informative features for the classification of Urdu news articles. The proposed methodology
using an ensemble of five popular ML classifiers with soft voting obtained 8th rank in the shared task
with an average macro F1-score of 0.592.
Keywords
UrduFake, Feature Engineering, Feature Selection, Machine Learning
1. Introduction
Fake news on social media is a common issue nowadays with the recent events in the world
such as the outbreak of Covid-19, return of Taliban to Afghanistan, etc. These situations and
the attributes of social media in being anonymous provide a lot of opportunities to content
polluters to share false information on social media networks [1, 2]. The task of fake news
detection is to identify the news articles that contain legit contents and distinguish them from
false information. Moreover, fake news detection is a significant issue since fake news might be
Forum for Information Retrieval Evaluation, December 13-17, 2021, India
" frs_b@yahoo.com (F. Balouchzahi); hlsrekha@gmail.com (H. L. Shashirekha); sidorov@cic.ipn.mx (G. Sidorov)
~ https://sites.google.com/view/fazlfrs/home (F. Balouchzahi);
https://mangaloreuniversity.ac.in/dr-h-l-shashirekha (H. L. Shashirekha); https://www.cic.ipn.mx/~sidorov/
(G. Sidorov)
0000-0003-1937-3475 (F. Balouchzahi)
© 2021 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
CEUR
Workshop
Proceedings
http://ceur-ws.org
ISSN 1613-0073 CEUR Workshop Proceedings (CEUR-WS.org)
a use-case to manipulate peoples’ mind towards a country, community, religion or any other
objective for political purposes or induction to a war, etc. [3, 4]. Detecting fake news manually
is totally ruled out due to the increase in the number of social media users, increasing volume
of fake news and the speed at which the fake news content is being generated and spread on
social media [5]. Therefore, developing automatic tools and models to identify the real news
from fake ones have become the primary requirements of the day to prevent the dissemination
of false information quickly and efficiently [6].
Fake news detection is a remarkable issue in many languages other than English [7]. Several
studies, shared tasks and workshops are being conducted to tackle the fake news detection
and profiling in high-resource languages such as English, Spanish, and German. However,
very few works have explored issues related to fake news detection in low-resource languages.
One among them is UrduFake 20211 [8, 9] - a shared task at Forum for Information Retrieval
Evaluation (FIRE) 2021 organized in continuation with the previous shared task - UrduFake
2020 [6] to promote fake news detection in Urdu language. The goal of this shared task is to
classify Urdu news articles into one of the two classes: Fake and Real.
To detect fake news in Urdu language, we - team MUCIC, present a feature engineering-
based ML model to investigate the effectiveness of ensembled FS algorithms on traditional ML
classifiers using char and word n-grams features. The objective of the methodology is to keep
the features and model simple and focus on the FS part.
The rest of paper is organized as follows: Section 2 gives a summary of the works submitted
to UrduFake 2020 followed by the description of Methodology in Section 3. The Experiments
and results are mentioned in Section 4 and the paper concludes in Section 5.
2. Related Work
Fake news detection in a resource-poor language like Urdu is a real challenge due to lack
of resources. UrduFake 20202 [6] is a shared task in FIRE 20203 that aims at identifying fake
news from real ones which can be modeled as binary Text Classification (TC). The shared task
participants were provided with a Train and Test set and the statistics of the datasets is given in
Table 1. Further, the models submitted by the participants were evaluated and ranked based
on macro F1-score. Several researchers had submitted their models to this shared task and the
description of the models which exhibited good performance are given below:
Amjad et al. [6] developed “Bend the truth” - a benchmarked dataset for fake news detection in
Urdu language and its evaluation [10]. The news items collected from Business, Health, Showbiz,
Sports, and Technology domains consisted of 500 real and 400 fake news. They collected real
news from different reliable sources such as BBC Urdu News, CNN Urdu, Roznama Dunya, Voice
of America, Mashriq News, etc. and the fake news were manually generated by hired journalists
from various news agencies such as Express news and Dawn news from Pakistan. The authors
also experimented several baselines with various weighting schemes. LR classifier trained on
Term Frequency–Inverse Document Frequency (TF-IDF) vectors generated from char bi-grams
1 https://www.urdufake2021.cicling.org/
2 https://www.urdufake2020.cicling.org/
3 http://fire.irsi.res.in/fire/2020/home
Table 1
Statistics of the Train and Test sets in UrduFake 2020
Categories
Dataset Total
Fake Real
Train set 400 500 900
Test set 150 250 400
exhibited the best performance obtaining 2nd rank with a macro F1-score of 0.889. Lin et al. [11]
represented text as word and char level units and then obtained sentence embedding in terms
of word and char level vectors using Robustly optimized Bidirectional Encoder Representations
from Transformers approach (RoBERTa) and Char Convolutional Neural Network (CharCNN)
respectively and fed them to softmax as a connecting layer to predict the authenticity of the
text. Utilizing label smoothing to enhance the performance of the model based on a smoothing
factor and the number of categories, this model secured 1st rank with a macro F1-score of 0.900.
The XLNet architecture benefits the generalized autoregressive method that utilizes autoen-
coding and autoregressive language modeling schemes [12]. Additionally, integration of the
recurrence mechanism and relative encoding scheme are the other advantages of XLNet. Khilji
et al. [13] fine-tuned XLNet [12] model with the maximum length of sequences to 512, and
used an architecture of six layers with four attention heads and an embedding dimension of
256. With this settings they obtained a macro F1-score of 0.823 and 2nd rank in the shared
task. Kumar et al. [14] proposed two models based on char n-grams in the range (1, 3) that
were transformed to Term Frequency — Inverse Document Frequency (TF-IDF) vectors. While
the first model is an ensemble of Random Forest (RF), Decision Tree (DT), and AdaBoost (AB)
with majority voting, the other model is a multi-layer dense Neural Network (NN) architecture
comprised of four dense layers containing 2,048, 512, 128, and 2-neurons with a dropout layer
after each dense layer. Using Rectified Linear unit (ReLu) activation function for the first three
dense layers and softmax activation function for the fourth layer, the multi-layer dense NN
model secured 3rd rank with a macro F1-score of 0.791.
Out of the three models experimented by Reddy et al. [15] the first two models are based
on bi-directional Gated Recurrent Unit (GRU) that employed Skipgram word embedding with
similar architecture and difference in only pooling layer. While max pooling and average pooling
were concatenated in one model, the other model employed only average pooling for the spatial
connections. However, in both the models the output of the pooling layer is fed to the sigmoid
layer for classification. The authors also experimented a multi-head transformer containing a
pipeline of vector generation along with a global max pooling layer and a hidden layer with a
dropout of 10% after each layer followed by softmax as final layer for classification. Among the
three models, the model using max pooling and average pooling exhibited the best performance
with a macro F1-score of 0.807. Balaji et al. [16] transformed word and char n-grams into TF-IDF
vectors along with pre-trained word embeddings, namely: Word2Vec, fastText, and Bidirectional
Encoder Representations from Transformers (BERT) trained on the traditional ML classifiers,
namely: Multilayer Perceptron (MLP), AB, ExtraTrees (ET), RF, Support Vector Machine (SVM),
and Gradient Boosting (GB). MLP classifier trained on fastText vectors obtained macro F1-score
of 0.788 and outperformed the other classifiers.
Figure 1: Overview of the proposed methodology
Balouchzahi et al. [17] experimented three learning approaches, namely: ML, Deep Learning
(DL), and Transfer Learning (TL). The ML model was an ensemble of LR, SVM, and Multinomial
Naïve Bayes (MNB) trained on the count vectors of the combination of word and char n-grams.
A BiLSTM architecture employing Skipgram word embedding generated from the training set
was used as DL model and for TL model Universal Language Model Fine-Tuning (ULMFiT)
architecture borrowed from Howard et al. [18] was used. ULMFiT model which makes use of a
pre-trained and publicly available Language Model4 fine-tunes the training set for classification.
The authors also submitted a majority voting of all the learning models to the shared task.
However, the best performance was exhibited by the ensemble of traditional ML classifiers with
a macro F1-score of 0.789.
3. Methodology
The proposed methodology contains two main steps: Feature Engineering (FE) and Model
construction and the strength of the methodology lies in the FE part. As the data is already
pre-processed only the punctuation if any are removed. Overview of the proposed methodology
is shown in Figure 1.
3.1. Feature Engineering
As the main objective of this work is to investigate the effectiveness of FS algorithms on
simple features, char and word n-grams are considered as features and are extracted separately.
Selecting the range of n-grams in an intelligent way could improve the performance of the
system. However, in this work, the range of n for char and word n-grams are set to (1, 3) and (2,
5) respectively, based on the results obtained by conducting various experiments.
The aim of FS algorithms is to reduce the number of features by selecting the relevant features
which may improve the performance of the model. An ensemble of two or more FS algorithms
will result in better features as compared to a single FS algorithm. However, ensembling may be
done in various ways. One way of ensembling is to consider the intersection of all the feature
sets obtained by the FS algorithms in the ensemble, ie., a feature is selected by an ensembled FS
4 https://github.com/anuragshas/nlp-for-urdu
Figure 2: Steps involved in selecting the features using ensembled FS algorithm
algorithm if and only if it is selected by all the FS algorithms. Algorithm 1 gives the ensembled
FS algorithm based on the intersection property of sets. The pictorial representation of the steps
involved in selecting the features using ensembled FS algorithm is shown in Figure 2.
Algorithm 1 Ensembled FS algorithm
1: procedure Ensembled_FS(features, FS) ▷ FS: feature selection algorithms
2: Selected_features = { }
3: for each FS algorithm do :
4: Compute the feature importance scores for all the features
5: Rank the features based on scores
6: Select the features with top 15,000 ranks and store in Feat_i
7: ▷ i for ith FS algorithm
8: for each feature f do:
9: if f in Feat_i and f in Feat_i+1 and ... and f in Feat_n then
10: Selected_features.add(f) ▷ intersection of all feature sets
11: else
12: Continue;
13:
return Selected_features
The FS algorithm explored by Shashirekha et al. [19] is an ensemble of three FS algorithms,
namely: Chi-square, Mutual Information Gain (MIG), and f_classif from sklearn library. This
ensembling algorithm is used in the proposed methodology to select the relevant features. The
char and word n-grams features are selected in two levels as follows:
• 30,000 top frequent features are selected from the given dataset and converted into TF-IDF
vectors (the number 30,000 is fixed based on the results obtained by conducting various
experiments)
• out of 30,000 top frequent features, the relevant features are selected by applying the
ensemble of FS algorithms explicitly
The number of n-grams features of both types at various stages are given in Table 2. Features
selected by the ensembled FS algorithm for the Train set is used to construct the models and for
that of Development (Dev.)/Test set are used to evaluate the models.
Table 2
Number of features in each stage
Feature All features Frequent features Selected features
Char n-grams 262,137 30,000 22,477
Word n-grams 355,030 30,000 22,616
Table 3
Classifiers and their parameter values
Classifier Parameters
LR Default parameters
Linear SVM Kernel="linear", probability= "True"
RFC N_estimators= 1000
hidden_layer_sizes= (150,100,50), max_iter=300,
MLP
activation = "relu", solver= "adam", random_state=1
max_depth= 20, n_estimators= 80, learning_rate= 0.1,
XGB colsample_bytree= 0.7, gamma=.01, reg_alpha=4,
objective= "multi:softmax", num_class= 2
3.2. Model Construction
To evaluate the effectiveness of the proposed FE algorithm, five popularly used traditional ML
classifiers, namely: Linear SVM (LSVM), LR, MLP, XGB, and RF are trained individually and
also as ensembled model with soft and hard voting. The individual and ensemble classifiers are
trained with and without FS to compare the effectiveness of FE module on various classifiers.
The classifiers and their parameters values used in this work are presented in Table 3.
4. Experiments and Results
UrduFake 2021 shared task organizers provided a dataset consisting of Train and Dev. set to
enable the participants to build and evaluate the models locally. However, for final evaluation,
Test set was provided without labels and the participating teams were supposed to predict the
labels of the Test set and submit the predictions to the organizers for evaluation and ranking
which was done based on average macro F1-score. The statistics of the datasets given in Table 4
shows that the datasets are highly imbalanced and it may negatively effect the performance of
the system.
As the number of runs that can be submitted by a participant was restricted to 4, four sets of
predictions were obtained on the Test set by applying the ensemble of five ML classifiers with
hard voting, soft voting, with FS and without FS (only frequent features) and were submitted to
the shared task organizers for evaluation and ranking.
A comparison of the performances of all the classifiers with and without FS (only frequent
features) for Dev. and Test set in terms of average macro F1-score is shown in Figures 3 and
4 respectively. The performances of many individual and ensembled classifiers with FS are
good compared to the ones without FS which justifies the effectiveness of FS algorithms. The
results also reveals that the LR classifiers with FS has given the best performance among all the
Table 4
Statistics of the datasets in UrduFake 2021
Dataset Fake Real Total
Train set 438 600 1,038
Dev. set 112 150 262
Test set 100 200 300
Figure 3: Performances on the Dev. set
experiments on the Test set with a macro F1-score of 0.617. However, the ensembled classifier
with soft voting and FS have obtained 8th rank with a macro F1 score of 0.592.
The comparison of the performances of the proposed model (along with the best results)
with the best performing teams of the shared task is given in Table 5. It can be observed that
the LR classifier with FS could have obtain 5th rank in the shared task if it was submitted to
the shared task (including baseline). Generally, the difference of less than 0.2 macro F1-score
between 1st and 10th ranks in the shared task shows very competitive performances among the
participating teams.
5. Conclusion and Future Work
This paper describes the methodology proposed by our team MUCIC for Urdu fake news
detection in UrduFake 2021 shared task. The main objective of the methodology lies in FE part
where three FS algorithms are ensembled to select the most informative char and word n-grams.
Effectiveness of the ensembled FS algorithm studied by using individual ML classifiers and their
ensemble as voting classifiers have shown promising results. Further, it is expected that other
feature sets such as sub-words, syllables and morphological features may significantly improve
Figure 4: Performances on the Test set
Table 5
Comparison of the best performing teams in UrduFake 2021
Fake Real Average
Team Name Accuracy
macro F1-score macro F1-score macro F1-score
Nayel 0.522 0.836 0.679 0.756
Abdullah-Khurem 0.530 0.797 0.663 0.716
Baseline 0.508 0.794 0.651 0.710
Hamed-Khurem 0.434 0.808 0.621 0.713
LR with FS
0.400 0.834 0.617 0.7400
(out of competition)
Muhammad Homayoun 0.485 0.738 0.611 0.653
Snehaan bhawal 0.384 0.837 0.610 0.743
Junaid 0.420 0.794 0.607 0.696
MUCIC
0.359 0.826 0.592 0.726
(soft voting with FS)
BERT4ever 0.438 0.746 0.592 0.650
the performance of the models. The future plans include exploring robust FE-based models
with various feature types and FS algorithms for TC in general in low resource languages such
as Urdu, Persian and Dravidian languages.
Acknowledgments
Team MUCIC sincerely appreciate the organizers for their efforts to conduct this shared task.
References
[1] H. L. Shashirekha, F. Balouchzahi, ULMFiT for Twitter Fake News Spreader Profiling, in:
Working Notes of CLEF 2020 - Conference and Labs of the Evaluation Forum, Thessaloniki,
Greece, September 22-25, 2020, volume 2696 of CEUR Workshop Proceedings, CEUR-WS.org,
2020. URL: http://ceur-ws.org/Vol-2696/paper_126.pdf.
[2] F. Balouchzahi, H. L. Shashirekha, G. Sidorov, MUCIC at CheckThat! 2021: FaDo-Fake News
Detection and Domain Identification using Transformers Ensembling, in: Proceedings of
the Working Notes of CLEF 2021 - Conference and Labs of the Evaluation Forum, Bucharest,
Romania, September 21st - to - 24th, 2021, volume 2936 of CEUR Workshop Proceedings,
CEUR-WS.org, 2021, pp. 455–464. URL: http://ceur-ws.org/Vol-2936/paper-35.pdf.
[3] M. Amjad, G. Sidorov, A. Zhila, A. Gelbukh, P. Rosso, UrduFake@ FIRE2020: Shared
Track on Fake News Identification in Urdu, in: FIRE 2020-Proceedings of the 12th Annual
Meeting of the Forum for Information Retrieval Evaluation, Association for Computing
Machinery, 2020, pp. 37–40.
[4] J.-P. Posadas-Durán, H. Gómez-Adorno, G. Sidorov, J. J. M. Escobar, Detection of Fake
News in a New Corpus for the Spanish Language, Journal of Intelligent & Fuzzy Systems
36 (2019) 4869–4876.
[5] N. Ashraf, S. Butt, G. Sidorov, A. F. Gelbukh, CIC at CheckThat! 2021: Fake News Detection
Using Machine Learning and Data Augmentation, in: Proceedings of the Working Notes of
CLEF 2021 - Conference and Labs of the Evaluation Forum, Bucharest, Romania, September
21st - to - 24th, 2021, volume 2936 of CEUR Workshop Proceedings, CEUR-WS.org, 2021, pp.
446–454. URL: http://ceur-ws.org/Vol-2936/paper-34.pdf.
[6] M. Amjad, G. Sidorov, A. Zhila, A. F. Gelbukh, P. Rosso, Overview of the Shared Task on
Fake News Detection in Urdu at FIRE 2020, in: Working Notes of FIRE 2020 - Forum for
Information Retrieval Evaluation, Hyderabad, India, December 16-20, 2020, volume 2826 of
CEUR Workshop Proceedings, CEUR-WS.org, 2020, pp. 434–446. URL: http://ceur-ws.org/
Vol-2826/T3-1.pdf.
[7] M. Amjad, G. Sidorov, A. Zhila, Data Augmentation using Machine Translation for Fake
News Detection in the Urdu Language, in: Proceedings of The 12th Language Resources
and Evaluation Conference, 2020, pp. 2537–2542.
[8] M. Amjad, B. Sabur, I. A. Hamza, Z. Alisa, S. Grigori, G. Alexander, Overview of the Shared
Task on Fake News Detection in Urdu at Fire 2021, in: Working Notes of FIRE 2021 - Forum
for Information Retrieval Evaluation, India, December, 2021, CEUR Workshop Proceedings,
CEUR-WS.org, 2021.
[9] M. Amjad, B. Sabur, I. A. Hamza, Z. Alisa, S. Grigori, G. Alexander, UrduFake@ FIRE2021:
Shared Track on Fake News Identification in Urdu., in: Working Notes of FIRE 2021 - Forum
for Information Retrieval Evaluation, India, December, 2021, In Forum for Information
Retrieval Evaluation, CEUR-WS.org, 2021.
[10] M. Amjad, G. Sidorov, A. Zhila, H. Gómez-Adorno, I. Voronkov, A. Gelbukh, "Bend the
Truth": Benchmark Dataset for Fake News Detection in Urdu Language and its Evaluation,
Journal of Intelligent & Fuzzy Systems 39 (2020) 2457–2469.
[11] N. Lin, S. Fu, S. Jiang, Fake News Detection in the Urdu Language using CharCNN-
RoBERTa, in: Working Notes of FIRE 2020 - Forum for Information Retrieval Evaluation,
Hyderabad, India, December 16-20, 2020, volume 2826 of CEUR Workshop Proceedings,
CEUR-WS.org, 2020, pp. 447–451. URL: http://ceur-ws.org/Vol-2826/T3-2.pdf.
[12] Z. Yang, Z. Dai, Y. Yang, J. G. Carbonell, R. Salakhutdinov, Q. V. Le, XLNet: Gen-
eralized Autoregressive Pretraining for Language Understanding, in: Advances in
Neural Information Processing Systems 32: Annual Conference on Neural Informa-
tion Processing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver, BC,
Canada, 2019, pp. 5754–5764. URL: https://proceedings.neurips.cc/paper/2019/hash/
dc6a7e655d7e5840e66733e9ee67cc69-Abstract.html.
[13] A. F. U. R. Khilji, S. R. Laskar, P. Pakray, S. Bandyopadhyay, Urdu Fake News Detection using
Generalized Autoregressors, in: Working Notes of FIRE 2020 - Forum for Information
Retrieval Evaluation, Hyderabad, India, December 16-20, 2020, volume 2826 of CEUR
Workshop Proceedings, CEUR-WS.org, 2020, pp. 452–457. URL: http://ceur-ws.org/Vol-2826/
T3-3.pdf.
[14] A. Kumar, S. Saumya, J. P. Singh, NITP-AI-NLP@ UrduFake-FIRE2020: Multi-layer Dense
Neural Network for Fake News Detection in Urdu News Articles., in: Working Notes of
FIRE 2020 - Forum for Information Retrieval Evaluation,Hyderabad, India, December 16-20,
2020, volume 2826 of CEUR Workshop Proceedings, CEUR-WS.org, 2020, pp. 458–463.
[15] S. M. Reddy, C. Suman, S. Saha, P. Bhattacharyya, A GRU-based Fake News Prediction
System: Working Notes for UrduFake-FIRE 2020, in: Working Notes of FIRE 2020 -
Forum for Information Retrieval Evaluation, Hyderabad, India, December 16-20, 2020,
volume 2826 of CEUR Workshop Proceedings, CEUR-WS.org, 2020, pp. 464–468. URL: http:
//ceur-ws.org/Vol-2826/T3-5.pdf.
[16] N. N. A. Balaji, B. Bharathi, SSNCSE_NLP@ Fake News Detection in the Urdu Language
(UrduFake) 2020, in: Working Notes of FIRE 2020 - Forum for Information Retrieval
Evaluation, Hyderabad, India, December 16-20, 2020, volume 2826 of CEUR Workshop
Proceedings, CEUR-WS.org, 2020, pp. 469–473.
[17] F. Balouchzahi, H. L. Shashirekha, Learning Models for Urdu Fake News Detection, in:
Working Notes of FIRE 2020 - Forum for Information Retrieval Evaluation, Hyderabad,
India, December 16-20, 2020, volume 2826 of CEUR Workshop Proceedings, CEUR-WS.org,
2020, pp. 474–479. URL: http://ceur-ws.org/Vol-2826/T3-7.pdf.
[18] J. Howard, S. Ruder, Universal Language Model Fine-tuning for Text Classification, in:
Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics
(Volume 1: Long Papers), 2018, pp. 328–339.
[19] H. L. Shashirekha, M. D. Anusha, N. S. Prakash, Ensemble Model for Profiling Fake News
Spreaders on Twitter, in: Working Notes of CLEF 2020 - Conference and Labs of the
Evaluation Forum, Thessaloniki, Greece, September 22-25, 2020, volume 2696 of CEUR
Workshop Proceedings, CEUR-WS.org, 2020. URL: http://ceur-ws.org/Vol-2696/paper_136.
pdf.