=Paper=
{{Paper
|id=Vol-3171/paper16
|storemode=property
|title=Evaluation of Vader and MultiLingual sentiment analyzers for opinion analytics using graphical illustrations
|pdfUrl=https://ceur-ws.org/Vol-3171/paper16.pdf
|volume=Vol-3171
|authors=Taiwo Olaleye,Peter Ugege,Olalekan Okewaler,Olatunji Akinade,Segun Akintunde,Chidi Akparanta
|dblpUrl=https://dblp.org/rec/conf/colins/OlaleyeUOAAA22
}}
==Evaluation of Vader and MultiLingual sentiment analyzers for opinion analytics using graphical illustrations==
https://ceur-ws.org/Vol-3171/paper16.pdf
Evaluation of Vader and MultiLingual sentiment analyzers for
opinion analytics using graphical illustrations
Taiwo Olaleye1, Peter Ugege2, Olalekan Okewale3, Olatunji Akinade4, Segun Akintunde5,
Chidi Akparanta6
1, 5
Computer Science Department, Federal University of Agriculture Abeokuta, Nigeria
2
Information & Communication Technology Centre, Forestry Research Institute of Nigeria
3
Directorate of Pan African University for Life & Earth Sciences Institute, University of Ibadan, Nigeria
4
Department of Cybersecurity, University of the West of England, Bristol
6
Computer Centre & Services, Federal College of Education Abeokuta, Nigeria
Abstract
Sentiment analysis have been prominently employed for opinion analytics across different use
cases which seeks to compute polarity values on opinions expressed in textual documents. The
Vader-based lexicon approach is widely deployed especially for tweet corpus with emoticon
elements. However, there is need to ascertain the peculiarities of the MultiLingual sentiment
analyzer alongside the Vader approach towards improving future studies in their use case
domain. The Twitter application programming interface was employed in this study to extract
public opinions in two corpus, on the Africa Cup of Nations matches involving Nigeria.
Experimental results shows no one-to-one mapping between the sentiment scores returned by
the two analyzers while the MultiLingual analyzer proves to be reputable for analyzing tweets
with shorter phrases. Tokens returned as the most weighted in the two corpus, as analyzed by
the two methodologies, likewise shows obvious contrasts in their weights. Two Nigerian
players were returned as prominent topics from the two corpus by the topic modelling phase
of the study while in MultiLingual analysis, local dialect tokens outweighs other English
unigrams, unlike in the Vader-based lexicon approach.
Keywords 1
Vader analysis, MultiLingual analysis , Football tweets, Opinion mining
1. Introduction
The popularity of footbal across the globe and internet access has provided a platform for cross-
fertilization of ideas and opinion moulding through social media platforms, as footbal remains the most
watched sport and played in over two hundred countries all over the world [1]. Moreover, the recent lift
of ban on Twitter by the Nigerian government on 13th January, 2022 came at a time the bi-annual
African Cup of Nations was starting. Young tweets consequently trooped to Twitter to kickstart their
support for the nation’s national team with different hashtags for different matches. The big opinion
corpus generated before, during, and after each group stage matches trends on Twitter for several hours
with divergent topical issues encapsulated in individual tweets. As a popular social microbloggin tool
employed by nationals and state actors for information dissemination and national integration, Twitter
serves as a village square for global, national, and sub-national digital town hall [2]. Opinions freely
expressed by users could be subjected to an opinion mining analytics by state actors to gauge the
national mood and probably influence state policies. Football being a capital intensive sport, and a tool
COLINS-2022: 6th International Conference on Computational Linguistics and Intelligent Systems, May 12–13, 2022, Gliwice, Poland.
EMAIL: agsobaolaleyetaiwo@gmail.com (T. Olaleye); pugege@yahoo.com (P. Ugege); okewale.akinbosoye@gmail.com (O. Okewale);
olatunji2.akinade@live.uwe.ac.uk (O. Akinade); seggzy2001@gmail.com (S. Akintunde); uzochidi2018@yahoo.com (C. Akparanta)
ORCID: 0000-0001-6222-7575 (T. Olaleye); 0000-0002-2722-0076 (P. Ugege); 0000-0003-2272-8660 (O. Okewale); 0000-0002-6126-6801
(S. Akintunde); 0000-0002-9493-1134 (C. Akparanta)
©️ 2022 Copyright for this paper by its authors.
Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
CEUR Workshop Proceedings (CEUR-WS.org)
for socioeconomic and national cohession, scrapping of opinions expressed in tweets for analytics
would serve stakeholders in different capacities [3]. Consequent upon the rude shock that trails the
suden exit of Nigeria from the competition, twitter has been agog with an upsurge of opinions on the
remote courses of the goal loss to the Tunisians despite a previous win against Guine Bisau. This study
therefore employs the instrumentality of Natural Language Processing to analyse shades of opinions
expressed on twitter on the two matches played by Nigeria. The Twitter Application Programming
Interface (API) is employed for the acquisition of tweets on the Nigeria match against Guine Bisau,
with a favourable outcome, and the match against Tunisia, with a loss. The dual corpus are subjected
to sentiment analysis with VADER and the Multilingual sentiment analyzers. Graphical representations
of the results are used for the overall analysis of the opinions expressed by Nigerians. The rest of the
paper is presented in the following format. Section II discusses related srudies while section III presents
materials and methods. Section IV discusses the results while the conclussion and recommendations
are presented in section V.
2. Related studies
Studies in the area of sentiment analysis, topic modelling, semantic extraction etc. are numerous in
the era of social media and better internet penetration. Employing same Natural Language Processing
use cases for football analysis is an emerging genre with global interest owing to the high
socioeconomic premium placed on football competitions across the globe. This section aims to discuss
similar studies in the area of football opinion mining studies, and sentiment analysis. In [4], an opinion
mining analytics aimed at discovering Covid-19 fear trigger was carried out on textual data using NLP.
The study’s topic modelling approach return hand washing as the major bigram that triggers fear of
Omicron variant of the pandemic, through the instrumentality of Vader-based lexison approach. The
work of [1] was particular about footballing opinion analytics together with prediction of event. A
perception model to understudy football players’ performances was the thrust of [5] through their online
presence. Feedbacks from blogs, editorials, articles, nespapers, and social medias were analyzed which
asigns ratings to each player. A football game analysis was conducted by [6] by proposing a framework
that predicts results of English Premier League through understudying of past games. A sentiment
analyser is incorporated from where a player could gauge the pulse of its fans over a match and likewise
has a tool that aids fans on betting. The analysis of women and men sentiments about a match was
implemented in the work of [7] and result shows certain similarities in the gender-corpus unexpectedly.
Men and women are seen to express similar emotional outbursts like anger, fear, etc. over Premier
league match outcomes. A sentiment analysis of tweets discussing football matches was conducted in
[8] to identify specific emotions expressed over football-related tweets. The domain-specific approach
of the study will further unravel the semantic meaning of expressed emotions and thereby design a
sentiment classifier capable of classifying expressed sentiments in football tweet conversations. The
prediction of wins and its spread I the Premier league was the main objective of [3] using sentiment
analysis. The aim is to identify if sentiments expressed in tweets over matches would be sufficient to
poredict future match outcomes and likewise discover if the magnitude of outcome would be correctly
predicted based on the degree of sentiments expressed. The study discovered that surge in positive
sentiments could net a payout of $3011.20 and also discover that the magnitude of positive sentiment
likewise correlates with a point spread. A lexicon-based sentiment analysis study was carried out by
[9]. The study tests the feasibility of lexicon-based analyzer for analysing football-based tweets. Over
10000 tweet corpus were acquired related to top-ten football matches for the purpose of the analysis.
Experiemntal result shows that sentiment of realistic sets with a proportion of 60% of same polarity can
be classified correctly with over 95% accuracy. Sentiment of world cup emotions to study theories of
emotions was the thrust of the study carried out in [10]. Contrary to consensus in sport economics, the
study revelaed that excitement relates to expression of negative emotions. In [11], a novel approach for
multilingual sentiment analysis was implemented on MultiLingual analyzer. Local language of Urdu,
English, and Roman-Urdu were analyzed with a novel dictionary on multilingual sentiments.
Experimental result shows extreme lexisons with high weights that are used to label the data returned
good performance. Sentiment analysis of Indonesian and French lexicons were analyzed by the
SentiSAIL tool in the work of [12]. The multilingual tool efficiently processed the languages with
reliable experimental results.
3. Materials & Methods
This section discusses the materials and methods employed in this study which is aimed at
unravelling emotions triggered by the outcome of football matches through the analysis of polarities
retumned by two sentiment analysers. The framework implemneted for the purpose of this study is as
captured in Figure 1.
3.1 Data acquisition
Twitter API was deployed for the acquisition of tweets through the instrumentality of hashtags,
which is a promising reference tool with prominent usage on twitter in [2]. Two tweet corpus was
captured for the purpose of this study including tweets discussing the outcome of the Nigerian match
against Tunisia and the tweets discussing the Nigerian match against Gambia. As represented on Table
1, a total of 141371 unigram of tweet corpus, representing 1000 tweet authors, was captured for the
Nigeria/Tunisia corpus while a total of 127072 unigram of tweets, representing 1000 authors, was
captured for the Nigeria/GuineBisau match.
3.2 Corpus preprocessing
This study employs the following corpus preprocessing techniques as recommended in [2], to return
a clean and tokenized unigram for the subsequent opinion mining analytics:
1. Conversion of the entire N_T and N_G corpus to lower case unigrams
2. Elimination of noises and white spaces from both corpus
3. A Regexp-based tokenization approach transforming entire corpuses into unigrams of useful tokens.
4. Filtering of entire corpus for the elimination of stop words like articles, prepositions, and
conjunctions that do not add to the emotion-based pattern recognition approach of the analysis
Table 1:
Tweet distribution across the two corpus
Corpus Corpus_ID Hashtag Corpus Unigram
size size
Nigeria/Tunisia N_T #NIGTUN 1000 141371
Nigeria/GuineBisau N_G #GNBNGA 1000 127072
3.3 Topic Modelling of N_T and N_G corpuses
Abstraction of subjects from textual corpus facilitated by word clusters and their frequency of
occurrence is the main thrust of topic modelling [13] as contained in each 1000 documents that makes
up each N_T and N_G respectively. It is a use case approach to NLP which unravels topical germane
subjects from corpus for information retrieval purposes and opinion mining studies. The topic
modelling is employed in this study to further unravel the most significant subject of discuss in each
corpus for an interpretative purpose. The Latent Semantic Indexing (LSI) of the Orange data mining
toolkit [4] is employed on both N_T and N_G. LSI returns topics with negative and positive keywords
with their corresponding weights. Positive weights signify words that are highly illustrative of the topic
while for negative weights, the topic rarely occur.
3.4 Sentiment Analysis using VADER-based Lexicon
Opinions expressed in textual corpus are categorized into positive, negative and neutral emotions
during sentiment analysis employing emotion polarities. The VADER-approach grades each unigram
contained each 1000 documents, as contained in each N_T and N_G, to compute a compound score for
each pair of 1000 documents. The methodology calculates the sentiment grades supposing emotion is
associated with the existence of definite phrases (for bi-gram), or words (for unigram) as in the case of
this study. Opinions are therefore assigned some sentiment values popularly referred to as lexicons.
Similar to the work of [14], the occurrence rate of each unigram in the aforementioned dictionary, as
used in each tweet document, would determine the calculations of its positive, negative or neutral
polarity. The compound score, which is computed from the polarity scores of positive, negative, and
neutral sentiment, is used in this study to determine the ultimate sentiment expressed in each 1000
documents of both N_T and N_G. It depicts collected account of the three polarity values and is
computed as
y (1)
y=
√y 2 + α
where y is the addition of valence values of component unigrams, and α is 15 (default value), which is
the regularization constant. The compound score is normalized between -1 and +1 that signifies the
positivity and negativity extent of each tweet opinion.
3.5 Sentiment Analysis using MultiLingual analyzer
The other sentiment analyzer employed in this study on the two corpus is the multilingual sentiment
analyzer. It is an RNN-based framework that is reputable for limited corpus [15]. The deep learning-
based method could employ any of the document embedders like word2vec, LoVe, doc2vec etc. for
input layers [16], and then contains the hidden layers, to the output layers, in its methodological
framework. The output could then be deployed for sentiment analysis, classification, entity extraction,
translations, topic modelling, etc. This approach has proven to return an impressive amount of
invariance to the traditional sentiment analysis systems like the Vader variant [16]. This is therefore
employed in this study on the acquired N_T and N_G corpus for further analysis.
Figure 1: Orange toolkit widgets for study framework
4. Result & Discussion
The outcome of the dual sentiment analysis using Vader-based lexicon and MultiLingual deep
sentiment analyzers on the Nigeria/Tunisia (N_T) and the Nigeria/GuineaBisau (N_G) corpus is
discussed in this section alongside the output of the LDA Topic modelling on both N_T and N_G. The
experiemntal result for Vader approach is sorted with the compound score from the smallest polarity to
the highest in order to isolate and cluster the output into their polarity sub-categories. The ten documents
with the most positive sentiments expression by Vader approach in the N_T is presented in the clustered
bar chart in Figure 2. The corresponding clustered bar chart for the ten most positive sentiment
expression as computed by the deep MultiLingual sentiment analyzer on the same N_T is as presented
in Figure 3. The Vader-based analysis output for the ten most negative sentiments expressed in N_T is
presented in Figure 4 while the Multilingual variant is presented in Figure 5. Analysis executed by
Vader approach on N_G for the ten most negatively expressed sentiments is presented in Figure 6 while
that of the MultiLingual counterpart is as presented in Figure 7. The positive sentimental versions for
both Vader and MultiLingual on N_G is presented in Figure 8 and Figure 9 respectively. Further
analysis is done on the ten most positive sentiment opinions as analysed by Vader and MultiLingual on
N_T and the result is presented in the word cloud of Figure 10 while that of the ten most negative
sentimantal opinion is captured Figure 11. An overall topic modelling by LSI on N_T and N_G
respectively is presented in Figure 12 to unravel topical tokens with the element of consensus in each
1000-sized corpus.
On the N_T, notwithstanding the 1:0 loss of Nigeria to Tunisia, some opinions on the match were
expressed with positive compound sentiments which may not be unexpected. However, there was no
one-to-one mapping between the ten tweet documents that makes each cluster of Vader and Multilingual
classes despite belonging to the same category, which is indicative of the sharp contrast between their
polarity computations. The same observation goes for the negative polarity category of both analyzers
on N_T. However, it can be observed that the tweet documents that makes each positive and negative
clusters for MultiLingual on N_T are majorly tweets with limited words unlike the Vader clusters with
longer phrases. Similar observations are common with the positive and negative clusters of both Vader
and MultiLingual on N_G where no one-to-one mapping exists between the clusters and the documents
that makes the MultiLingual clusters are of shorter phrases compared to the Vader counterpart. In
positive sentiment cluster MultiLingual on N_T, tokens of local dialets and or pidgin English like
‘wahala’, ‘won’ etc., meaning trouble and want, accrues higher weights with no such traces in the Vader
word cloud. The ‘good’, ‘best’, and ‘small’ tokens posses highest weights in the Vader cluster. In the
negative category, MultiLingual similarly returned pidgin English dialet tokens like ‘sha’, ‘e’, ‘no’ etc.,
as against the only local dialet token ‘pass’ noticed in the Vader negative cluster of N_T. The context
in which they are used uniquely identifies a proper English from the pidgin local version. The negative
clusters of N_T over Vader and MultiLingual however returned highly negative tokens as those with
highest weights which is indicative of the mood of tweet posters who expresses their sentiments at the
loss of the match. Tokens like ‘bad’, ‘winch’ (meaning withcraft in pidgin english), ‘cry’, ‘pikin’
(meaning child in pidgin English), ‘die’, and ‘kill’, ‘pain’, ‘sorry’, etc. all encapsulates the mood of the
tweet documents in the negative N_T cluster. Prominent tokens like ‘Okoye’ and ‘Moses’ are the most
significant topics returned by LSI on N_T and N_G respectively besides other commonplace topics like
‘nigeria’, ‘afcon’, ‘ngatun’, and ‘gnbnga’. Experimental result of the topic modelling shows that the
name of the Nigeria goal keeper, Okoye, is prominent in the 141371 unigram that makes up the N_T
while the name of the Nigeria striker, Moses, was the most prominent in the 127072 unigram as
contained in the N_G corpus.
Vader Positive on N_T
Hi @Mpraiz beautiful reflex with your…
D Nigerian CEOs,companies and business…
Happiness no dey last for naija...…
APPRECIATION TWEET…
Well done to the Super Eagles! You…
Good morning Nigeria. The best hand…
Nevertheless I'm really proud of our…
Reality check for Super Eagles 🦅 . To…
Your Feet are the first things people…
Sadiq Umar was terrific tonight. He…
0 0.2 0.4 0.6 0.8 1 1.2
Comp Neu Neg Pos
Figure 2: Clustered bar chart for ten most positive polarities by Vader on N_T
MultiLingual Positive on N_T
Tactics won #NGATUN
The proper Eagles won.…
The Super Eagle right now…
Well it was fun while it lasted......…
Nothing extraordinary a new coach will…
Super Eagles Played well. Not our day.…
Kudos to our super 👌 eagles…
Well done to the Super Eagles! You…
Wahala no too much like this 🦅 🦅 …
Super Eagles really led us all on. Mfs!!!…
0 5 10 15 20 25 30
Figure 3: Clustered bar chart for ten most positive polarities by MultiLingual on N_T
Vader Negative on N_T
Nigeria team loses as a team. The ball…
Not Nigerians changing their names and…
No be anything but na overhyping kill dos…
Now our lil annoying bothers (Ghana) will…
Super Eagles just annoy me anyhow .I…
Eguavoen should have changed the…
While we are at it, can we talk about how…
Ghanaians fighting aimlessly over a loss.…
Winch cry yesterday; Pikin die today.…
Didn’t ya see that the ball changed…
-1.5 -1 -0.5 0 0.5 1
Comp Neu Neg Pos
Figure 4: Clustered bar chart for ten most negative polarities by Vader on N_T
MultiLingual Negative on N_T
0
-5
-10
-15
-20
-25
-30
-35
Figure 5: Clustered bar chart for ten most negative polarities by MultiLingual on N_T
Vader Negative on N_G
Guinea Bissau said they would rather die…
CAF should check on Guinea Bissau…
But Wait I Thought These Guys Said…
Guinea Bissau -“we would rather die than…
"Over our dead body"…
Didn’t Guinea Bissau say that they would…
Guinea captain: I will rather die than for…
It’s A shame @Guinea_Bissau Squad lost…
Ghana is dead oooo🦅 🦅 😂😂😂. This poor…
Guinea say them go murder us for pitch ..…
-1.2 -1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8
Comp Neu Neg Pos
Figure 6: Clustered bar chart for ten most negative polarities by Vader on N_G
MultiLingual Negative on N_G
0
-5
-10
-15
-20
-25
-30
Figure 7: Clustered bar chart for ten most negative polarities by MultiLingual on N_G
Vader Positve on N_G
With the 2-0 win over Guinea-Bissau, the…
Today issa a good day…
After we have won the AFCON…
Convincing win from the Super Eagles…
My baby Maduka Okoye didn’t play today,…
Jamilu Collins played really well…
Super eagles won 🦅 …
@samvicas @McDrMoore a good striker…
Congratulation @NGSuperEagles 🦅 🦅 🦅 …
🦅 🇳🇬 Nigeria Super Eagles ‘Moses Simon’…
0 0.2 0.4 0.6 0.8 1 1.2
Comp Neu Neg Pos
Figure 8: Clustered bar chart for ten most positive polarities by Vader on N_G
MultiLingual Positive on N_G
Best caption wins 😂 #GNBNGA…
Sweet sweet victory #TeamNigeria…
The unbeatable super eagles 🇳🇬🇳🇬💪…
A great job well done #GNBNGA
Beautiful football 🇳🇬#GNBNGA
Congratulations to the super team!!🎉🇳🇬…
#GNBNGA nice performance
Good game Super Eagles. Impressive!…
This Ajayi is a decent defender, he matches…
Congratulations to Super Eagles 🦅 🇳🇬🇳🇬🇳🇬…
0 5 10 15 20 25 30 35
Figure 9: Clustered bar chart for ten most positive polarities by MultiLingual on N_G
(a) (b)
Figure 10: Word cloud of ten most positive N_T tokens for (a) Vader and (b) MultiLingual
(a) (b)
Figure 11: Word cloud of ten most negative N_T tokens for (a) Vader and (b) MultiLingual
(a)
(b)
Figure 12: Result of five prominent topics for (a) N_T and (b) N_G respectively
5. Conclusion & Recommendation
This study evaluatyes the performances of the Vader-based lexicon sentiment analyzer and the
MultiLingual sentiment analyzers through an opinion mining analytics of football tweets. Two tweet
corpus of 1000 tweet each, comprising of 141371 and 127072 from the hashtag #NGRTUN and
#GNBNGA respectively. Experiemntal results of sentiment analysis and topic modelling reveals that
while Vader-based approach targets tweets of longer phrases, MultiLingual analyzer works with tweets
of shorter phrases. The MultiLingual analyzer likewise returned unigram of local dialets than the Vader
approachin the word cloud analysis. Two players were returned as the most discussed topics in the two
corpus by LSI topic modelling, including Okoye and Moses, who might have played significant roles
in the two matches where Nigeria lost the N_T match and won the N_G match. Further studies could
expand the scope of the textual corpus above 1000 to ascertain if MultiLingual would retain its penchant
for short phrases.
6. Acknowledgement
The authors appreciates the painstaking efforts of the reviwers which made this work turn out better.
7. References
[1] N. J. D. Souza, H. N. Samrudh, S. Gautham, B. U. S. Bhat and N. Nagarathna, "Football Game
Analysis and Prediction," in Soft Computing and Signal Processing, 2021.
[2] T. Olaleye, T. Arogundade, A. Abayomi-Alli and K. Adesemowo, "An Ensemble Predictive
Analytics of COVID-19 Infodemic Tweets using bag of words," in Data Science for COVID-19,
Singapore, Elsevier, 2021.
[3] R. P. Schumaker, A. T. Jarmoszko and C. S. L. Jr., "Predicting wins and spread in the Premier
League using a sentiment analysis of twitter," Decision Support Systems, pp. 76-84, 2016.
[4] T. O. Olaleye, S. M. Akintunde, C. Akparanta, T. A. Avovome, O. F. Oluyen and A. O. Akparanta,
"Opinion Mining Analytics for Spotting Omicron Fear-Stimuli Using REPTree Classifier and Natural
Language Processing," International Journal for Research in Applied Science & Engineering
Technology, pp. 995-1005, 2022.
[5] Z. Kabir, "Perception Model to Analyze Football Players' Performances," PACIS, 2019.
[6] n. Souza, H. Samrudh, S. Gautham, B. S. Bhat and N. Nagarathna, "Football Game Analysis and
Prediction," in Soft Computing and Signal Processing , 2021.
[7] M. B. Babac and V. Podobnik, "A sentiment analysis of who participates, how and why, at social
media sport websites: how differently men and women write about football," Online Information
Review, 2016.
[8] S. Aloufi and A. E. Saddik, "Sentiment identification in football-specific tweets," IEEE Access ,
2018.
[9] F. Wunderlich and D. Memmert, "Innovative approaches in sport science-lexicon-based sentiment
analysis as a tool to analyze sports-related Twitter communication," Applied Sciences, 2020.
[10] G. M. Lucas, J. Gratch, N. Malandrakis, E. Szablowski and E. Eessler, "GOALLL:: Using
sentiment in the world cup to explore theories of emotion," Image and Vision Computing, pp. 58-65,
2017.
[11] M. Asif, A. Ishtiaq, H. Ahmad, H. Aljuaid and J. Shah, "Sentiment analysis of extremism in social
media from textual information," Telematics and Informatics, 2020.
[12] G. Shalunts, G. Backfried and H. S. Alam, "Sentiment analysis in Indonesian and French by
SentiSAIL," in 017 Asia-Pacific Signal and Information Processing Association Annual Summit and
Conference, 2017.
[13] T. Sherstinova, O. Mitrofanova, T. Skrebtsova, E. Zamiraylova and M. Kirina, "Topic Modelling
with NMF vs. Expert Topic Annotation: The Case Study of Russian Fiction," in Mexican International
Confeence on Artificial Intelligence, 2020.
[14] H. Hota, D. Sharma and N. Verma, "Lexicon-based sentiment analysis using Twitter data: a case
of COVID-19 outbreak in India and abroad," in Data Science for COVID-19, London, Elsevier, 2021,
pp. 275-293.
[15] E. F. Can, A. Ezen-Can and F. Can, "Multilingual Sentiment Analysis: An RNN-Based Framework
for Limited Data," arXiv:1806.04511, 2018.
[16] P. Ghaffari, "Leveraging Deep Learning for Multilingual Sentiment Analysis," LinkedIn, 2016.