The Covid-19 epidemic situation reflected significant
changes in many dimensions, both in human behavior
consumption and service behavior. The SARS-Cov-2 known
as Covid-19 has spread around the world and it’s also
designation as a worldwide pandemic by the World Health
Organization in March 2020 [1]. While the world is struggling
to handle with Covid-19, the number of infected patients has
continuingly increased. Thailand has changed since Covid-19
and people are worried about the spread of this situation.
Information is important to be aware of the events in society.
People tend to follow the news through various media,
including real news, fake news, or current news. Especially in
this era of COVID-19, information has a huge impact on
people's emotions. Therefore, most people have anxiety,
paranoia, and fear that they or close people are infected which
results in confusion in information. Twitter is one of the
popular social media platforms that aims to provide users with
the ability to comment in short texts up to 140 characters.
Twitter presents the trend that people talk about or what
trending is on Twitter right now. On twitter message, a
hashtag is used to play critical roles in recent social
movements such as #election, #Covid-19, and etc. It is a word
or sentence that has a "#" preceding it. This is a form of
metadata tag that is widely used in social media. Hashtags
have played the important role in conversation. There was a
discussion to exchange comments and hashtags in Thailand
are used in a variety of ways.
Opinion mining is the science of gathering opinions from
multiple messages on a particular subject to analyze opinions.
It is often analyzed as positive, negative, or neutral.
Information extraction and sentiment analysis has been
broadly acknowledged as one of the first stages in the natural
language processing [2], [3]. This research is aimed to classify
the textual information on the social media platforms like
Twitter. The significant approaches, like K-NN, Naïve Bayes,
Decision Tree, Random Forest and Support Vector Machine
(SVM), were used to information extraction process and then
evaluated with the F1 score accuracy of each algorithm. In the
second process, Bi-directional GRU, one of the deep learning
method, was applied to use for sentiment analysis task.
Experimental result may be in charge of helping to the
development of public health properly.
This section presented a literature review of relevant
researches for exploration an overview of current knowledge
of sentiment analysis. Tweets from twitter [3] were classified
into positive, negative, and neutral. Dusmanu et al. [4]
applied argument mining methods to classify arguments on
Twitter from actual facts. Vaccine-related tweets were
analyzed and the results showed the number and the opinion
polarity of tweets in neutral 60%, 23% against vaccination,
and 17% in favor of vaccination [5].
Naïve Bayes model was implemented to analyze
sentiments towards COVID-19 with Twitter datasets in
English and Filipino language and the algorithm supports to
classify tweets by using Rapid Miner [6]. Machine learning
algorithms and lexicon-based approaches were proposed to
sentiment word detection and POS tagging [7]. According to
Tang, Kay and He [8], Naive Bayes (NB), and Support
Vector Machine (SVM) were used to Text Classification. To
analyze reliability, Naïve Bayes was adopted to identify the
untrusted content on Twitter [9]. Deep learning based on
Machine Learning Algorithms for Natural Language Processing
Tasks: A Case of COVID-19 Twitter data (Thailand)
1KUNYANUTH KULARBPHETTONG, 1RUJIJAN VICHIVANIVES, 2PANNAWAT
KANJANAPRAKARN, 2KANYARAT BUSSABAN, 2JARUWAN CHUTRTONG,
3NAREENART RUKSUNTORN
1Computer Science Program Suan Sunandha Rajabhat University Bangkok, THAILAND
2Faculty of Science and Technology Suan Sunandha Rajabhat University Bangkok, THAILAND
3Robotics Engineering program Faculty of Industrial Technology Suan Sunandha Rajabhat University
Bangkok, THAILAND
Abstract: This paper presents the use of natural language processing for the problem of information extraction
and sentiment analysis. The dataset is from Twitter that has the information of people mentioning about COVID-
19, this study has two tasks: (i) classification approach for information extraction task and (ii) deep learning
approach for sentiment analysis task. In information extraction task, the data was gathered from twitter that
related to COVID-19 information, and the sequence labelling method applied to classify text before giving it to
classification algorithms (K-NN, Naïve Bayes, Decision Tree, Random Forest, and SVM). In sentiment analysis
task, data was classified by convert the word into index and using word embedding, then to process deep
learning algorithm (Bi-directional GRU). The accuracy of two tasks are 98% and 79% respectively.
Keywords: COVID-19, KNN,Deep learning, Random Forest, Bi-directional GRU
Received: March 24, 2022. Revised: October 18, 2022. Accepted: November 21, 2022. Published: December 31, 2022.
1. Introduction
2. Literature Reviews
International Journal on Applied Physics and Engineering
DOI: 10.37394/232030.2022.1.5
Kunyanuth Kularbphettong et al.
E-ISSN: 2945-0489
31
Volume 1, 2022
LSTM, GRU, and CNN and feature-based methods were
combined to financial sentiment analysis [10]. Contextual
deep learning was applied to analyze in sentiment analysis
involves categorizing subjective opinions from text, audio,
and video sources [11].
This section describes the relevant approaches using conduct
this research.
The scope of this study is Thailand and the data was
considered news about COVID-19 in Thailand. In figure 1,
data was collected almost 600,000 tweets by using Tweepy (a
python library) [12] and then processed the raw data
(Unstructured data) to be data that is in the form of an
appropriate structure (structure data). The data was pre-
processed by cleaning and tokenization text using NLTK
library.
Fig. 1. A sample of data retrieved from the web page application
The task consists in classifying a tweets as containing
report information of coronavirus focused on particular tweets
patterns like “total 42 cases” or “500 total deaths” of their
sources. The five algorithms of classification including K-NN,
Naïve Bays, Decision Tree, Random Forest and Support
Vector Machine were used to classify tweets and evaluate the
results.
The labelling is the next process from pre-processing and
700 tweets were selected to find the amount of people who
affected in this pandemic. Then the data was annotated the
numbers that occur in text. This would allow us to understand
and make it easier to train the algorithms. These numbers are
annotated as “1” if it follows by “total cases”, or we annotated
as “2” if it follows by “total deaths”, if it not fits the above
conditions then we annotated as “0” (see example (a) and (b)
below).
(a) Text: “iran reports 3 new cases bringing total
confirmed cases 52 total deaths.”
Tag: “[0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 0, 0]”
(b) Text: “total deaths 75 total cases 100.”
Tag: “[0, 0, 2, 0, 0, 1]”
After manual labelling, the data set was prepared to be
processed with algorithms by transform “Text” to Python
dictionary by having key as the following.
“word” : The word itself.
“postag” : The part of speech of the word.
“nextword” : The next word after the word itself.
“nextwordtag” : The next part of speech tag of next word.
“previousword” : The previous word before the word
itself.
“previoustag” : The previous part of speech tag before the
word itself.
The data was spited it to 70:30 proportion training and
testing and transform Python dictionary to vector by using
DictVectorizer from Scikit-learn library [13].
This section presented the results of this research. K-NN,
Naïve Bays, Decision Tree, Random Forest and Support
Vector Machine were used to extract information and the
results were shown in table I and figure 2and 3.
TABLE I. RESULTS FOR INFORMATION EXTRACTION TASK
Precision
Recall
K-NN
0.93
0.91
0.92
Naïve Bayes
0.87
0.87
0.87
Decision Tree
0.93
0.90
0.92
Random Forest
0.95
0.91
0.93
SVM
0.94
0.91
0.93
Fig. 2. Results of Accuracy Scores
Fig. 3. Results of F1 Scores
3. Methodology
3.1 Data Set and Data Preparation
3.2 Information Extraction
4. Results
International Journal on Applied Physics and Engineering
DOI: 10.37394/232030.2022.1.5
Kunyanuth Kularbphettong et al.
E-ISSN: 2945-0489
32
Volume 1, 2022
When considered in each algorithm, the results were as
follows:
(a) KNN has an overall accuracy of 98%, a classification
precision in infected cases (1) 94% and precision in
classifying deaths (2) 93% as shown in figure 4.
Fig. 4. Results of KNN
(b) Naïve Bays has an overall accuracy of 97%, a
classification precision in infected cases (1) 85% and
precision in classifying deaths (2) 90% as shown in figure 5.
Fig. 5. Results of Naïve Bays
(c) Decision Tree has an overall accuracy of 98%, a
classification precision in infected cases (1) 94% and
precision in classifying deaths (2) 93% as shown in figure 6.
Fig. 6. Results of Decision Tree
(d) Random Forest has an overall accuracy of 98%, a
classification precision in infected cases (1) 95% and
precision in classifying deaths (2) 95% as shown in figure 7.
Fig. 7. Results of Random Forest
(e) Support Vector Machine has an overall accuracy of 98%,
a classification precision in infected cases (1) 94% and
precision in classifying deaths (2) 95% as shown in figure 8.
Fig. 8. Results of Support Vector Machine
From the previous results, the RF (Random Forest) algorithm
has a higher score than the other algorithms and Naive Bayes
has the lowest score. Therefore, this framework choose
Random Forest in the next process.
Bi-Directional GR, one of the Deep Learning approaches,
have used to experiment with modifying Word Embedding
by choosing Covid Word Embedding and English Word
Embedding, which gives accurate results as presented in table
2 and 3.
TABLE II. RESULTS OF THE TEST SET OF ENGLISH (COVID-19) WORD
EMBEDDING
Polarity
Precision
Recall
F1
POSITIVE
0.77
0.78
0.78
NEGATIVE
0.78
0.77
0.77
TABLE III. RESULTS OF THE TEST ENGLISH WORD EMBEDDING
Polarity
Precision
Recall
F1
POSITIVE
0.80
0.77
0.79
NEGATIVE
0.78
0.81
0.79
Figure 9 shows the construction process of the Bi-directional
GRU sentiment analysis classification model and two pre-
trained word embedding was generated by Fast-Text. First
word embedding is plain English text with no related to any
field, and the other is word embedding that related to Covid-
19.
Fig. 9. Bi-directional gated recurrent neural networks (GRU) sentiment
analysis model
International Journal on Applied Physics and Engineering
DOI: 10.37394/232030.2022.1.5
Kunyanuth Kularbphettong et al.
E-ISSN: 2945-0489
33
Volume 1, 2022
From table 4, it shows the English word embedding has better
accuracy than Covid word embedding, because our dataset
(Kaggle) that we use to train is not related to Covid fields.
Also, if we use the English word embedding in a real-time
Tweets about Covid, it will significantly decreased the
accuracy as well.
TABLE IV. RESULTS FOR INFORMATION EXTRACTION TASK
Word Embedding
Accuracy
English (Covid-19)
0.776
English
0.791
Fig. 10. Results of information Extraction
This study investigated information extraction and
sentiment analysis on Twitter data. These tasks are
particularly relevant when applied to social media data and the
Covid19 global pandemic. The issue of information extraction
on Twitter is we are labeling the data by manually unlike
sentiment analysis that is Kaggle dataset. Thus, the dataset on
information extraction is limited (700 tweets) not
comprehensive to the other report pattern which give us
limited result and accuracy. In future work, we will focus on
extending and increasing the datasets of information
extraction by augmentation method, and exploring more on
sentiment analysis dataset in order to have more reliability in
real-time use.
The authors express their sincere appreciation to Suan
Sunandha Rajabhat University for financial support of the
study.
[1] K. Chong Ng Kee Kwong, P. R. Mehta, G. Shukla, and A. R. Mehta,
“COVID-19, SARS and MERS: A neurological perspective,” Journal
of Clinical Neuroscience, May 2020. [Online]. Available:
http://www.sciencedirect.com/science/article/pii/S0967586820311851
[2] Ravi K., Ravi V., A survey of opinion mining and sentiment analysis:
Tasks, approaches and applications, Knowledge-Based Systems (89)
(2017), pp. 14-46
[3] Kunyanuth Kularbphettong, The awareness of environment
conservation based on opinion data mining from social media,
International Journal of GEOMATE, Sept., 2019 Vol.17, Issue 61, pp.
74 – 79
[4] Mihai Dusmanu, Elena Cabrio, and Serena Villata. Argument mining
on twitter: Arguments, facts and sources. In EMNLP, pages 2317–
2322, 2017
[5] Lara Tavoschi, Filippo Quattrone, Eleonora D’Andrea, Pietro
Ducange, Marco Vabanesi, Francesco Marcelloni & Pier Luigi Lopalco
(2020) Twitter as a sentinel tool to monitor public opinion on
vaccination: an opinion mining analysis from September 2016 to
August 2017 in Italy, Human Vaccines & Immunotherapeutics, 16:5,
1062-1069, DOI: 10.1080/21645515.2020.1714311
[6] Villavicencio, C.; Macrohon, J.J.; Inbaraj, X.A.; Jeng, J.-H.; Hsieh, J.-
G. Twitter Sentiment Analysis towards COVID-19 Vaccines in the
Philippines Using Naïve Bayes. Information 2021, 12, 204. https://
doi.org/10.3390/info12050204
[7] Park S, Kim Y. 2016. Building thesaurus lexicon using dictionary-
based approach for sentiment classification. In: 2016 IEEE 14th
International Conference on Software Engineering Research,
Management and Applications (SERA). Piscataway: IEEE, 39–44.
[8] Tang B, Kay S, He H. 2016. Toward optimal feature selection in naive
bayes for text categorization. IEEE Transactions on Knowledge and
Data Engineering 28(9):2508–2521 DOI
10.1109/TKDE.2016.2563436.
[9] M. AlRubaian, M. Al-Qurishi, M. Al-Rakhami, S. M. M. Rahman, and
A. Alamri, A Multistage Credibility Analysis Model for Microblogs,
presented at the Proceedings of the 2015 IEEE/ACM International
Conference on Advances in Social Networks Analysis and Mining
2015, Paris, France, 2015
[10] Akhtar MS, Kumar A, Ghosal D, Ekbal A, Bhattacharyya P. 2017. A
multilayer perceptron based ensemble technique for fine-grained
financial sentiment analysis. In: Proceedings of the 2017 Conference
on Empirical Methods in Natural Language Processing. 540–546.
[11] [Adeel A, Gogate M, Hussain A. Contextual deep learning-based
audio-visual switching for speech enhancement in real-world
environments. Information Fusion 2020 Jul;59:163-170. [CrossRef]
[12] [Tweepy G.e.(2020),Retrieved 2021, from Tweepy:
https://www.tweepy.org/
[13] DictVectorizer, Retrieved 2021, from scikit-learn.org: https://scikit -
learn.org/stable/modules/generated/sklearn.feature_extraction.DictVe
ctorizer.htmlAuthor No.1, Author No 2 Onward, “Paper Title Here”,
Proceedings of xxx Conference or Journal (ABCD), Institution name
(Country), February 21-23, year, pp. 626-632.
5. Conclusion
Acknowledgment
References
Creative Commons Attribution License 4.0
(Attribution 4.0 International, CC BY 4.0)
This article is published under the terms of the Creative
Commons Attribution License 4.0
https://creativecommons.org/licenses/by/4.0/deed.en_US
International Journal on Applied Physics and Engineering
DOI: 10.37394/232030.2022.1.5
Kunyanuth Kularbphettong et al.
E-ISSN: 2945-0489
34
Volume 1, 2022