International Journal of Electrical and Computer Engineering (IJECE)

Vol. 12, No. 5, October 2022, pp. 5562~5572

ISSN: 2088-8708, DOI: 10.11591/ijece.v12i5.pp5562-5572  5562

Sentiment analysis on Bangla conversation using machine

learning approach

Mahmudul Hassan1, Shahriar Shakil1, Nazmun Nessa Moon1, Mohammad Monirul Islam1,
Refath Ara Hossain1, Asma Mariam1, Fernaz Narin Nur2
1
Department of Computer Science and Engineering, Faculty of Science and Information Technology, Daffodil International University,
Dhaka, Bangladesh
2
Department of Computer Science and Engineering, Faculty of Science and Information Technology, Notre Dame University,
Dhaka, Bangladesh

Article Info ABSTRACT

Article history: Nowadays, online communication is more convenient and popular than face-
to-face conversation. Therefore, people prefer online communication over
Received Jun 10, 2021 face-to-face meetings. Enormous people use online chatting systems to
Revised May 23, 2022 speak with their loved ones at any given time throughout the world. People
Accepted Jun 20, 2022 create massive quantities of conversation every second because of their
online engagement. People's feelings during the conversation period can be
gleaned as useful information from these conversations. Text analysis and
Keywords: conclusion of any material as summarization can be done using sentiment
analysis by natural language processing. The use of communication for
Accuracy rate customer service portals in various e-commerce platforms and crime
Detection approach investigations based on digital evidence is increasing the need for sentiment
Natural language processing analysis of a conversation. Other languages, such as English, have well-
Sentiment analysis developed libraries and resources for natural language processing, yet there
Support vector machine are few studies conducted on Bangla. It is more challenging to extract
Tokenizer sentiments from Bangla conversational data due to the language's
grammatical complexity. As a result, it opens vast study opportunities. So,
support vector machine, multinomial naïve Bayes, k-nearest neighbors,
logistic regression, decision tree, and random forest was used. From the
dataset, extracted information was labeled as positive and negative.
This is an open access article under the CC BY-SA license.

Corresponding Author:
Nazmun Nessa Moon
Department of Computer Science and Engineering, Daffodil International University
Dhaka 1207, Bangladesh
Email: [email protected]

1. INTRODUCTION
People have conversations in their daily life. People express their feelings and opinions in their
conversations. These feelings and opinions can be categorized into sad, anger, happy, worried, disgusted,
frightened, complement, motivation, suggestions, and neutral [1]. To detect subjective information such as
opinions, attitudes, and feelings expressed in text Sentiment analysis or opinion mining aims to use
automated tools [2]. In our research work we merged them into two main categories of positive and negative
[3]. Sentiment analysis can be done by capturing both semantic and sentiment similarities among words [4].
Our model can identify whether a part of any conversation is positive or negative. These two categories
expose the sentiment of the people who said it. Analyzing sentiment from people’s speech is a tough job
because in a single sentence people can express various types of sentiment at the same time. Only the people
who listen to it, can understand the sentiment properly. Our proposed model can extract sentiment from

Journal homepage: http://ijece.iaescore.com

Int J Elec & Comp Eng ISSN: 2088-8708  5563

people’s conversation with a closer accuracy of real life. In this research work we proposed a model that can
extract sentiment from conversation as positive or negative sentiment. To pursue that we split our dataset into
80:20 ratio. For training purposes, we used 80% data and for testing purposes we used 20% data. It helps to
increase the accuracy of the model. Based on the training dataset the accuracy of the model fully depends on
the training dataset. We have used some techniques such as changing the parameters of machine learning
models to get more accurate results. We achieved about 86% accuracy on the support vector machine. Rest of
the algorithms perform closely to the highest accuracy.

2. LITERATURE REVIEW
Extracting sentiment from Bangla conversational data is a method for determining if a conversation
is positive or negative. Bhowmik et al. [5] developed deep learning models for Sentiment analysis on Bangla
text using an extended lexical data set. They employed the rule-based Bangla text sentiment score system to
extract polarity from large texts. These polarities, along with the pre-processed text, are then used as training
samples by the neural network. The pre-processed texts are displayed as a vectorization of words derived
from pre-trained word embedding models with various word counts. A Word2Vec matrix containing the top
highest probability word is used as a weighted matrix on the embedding layer to fit the deep learning models.
This paper also includes a thorough examination of selective deep learning models, as well as some fine-
tuning. Their proposed hierarchical approach was accurate to the tune of 78.52 percent, 80.82 percent, and
84.18 percent, respectively. According to Aurpa et al. [6] certain items, such as threats and sexual
harassment, were more accessible than traditional media. Harassment, vulgarity, personal assaults, and
bullying can all occur because of extremely toxic internet content. Bangla's use of Facebook has risen in
recent years due to its status as the world's seventh most spoken language. The use of offensive comments in
Bangla on Facebook has also grown significantly, but there is little research on the subject. They focus on
recognizing abusive Bangla language remarks on social media (Facebook) that can be filtered out in the early
phases of social media attachment in this study. To classify hostile comments quickly and accurately,
transformer-based deep neural network models were used. They employed pre-training language
architectures bidirectional encoder representations from transformers (BERT) and efficiency learning an
encoder that accurately classifies token replacements (ELECTRA). The average accuracy, precision, recall,
and f1-score were used to assess the proposed models. The results have revealed that our BERT and
ELECTRA architectures are performing admirably, with test accuracy of 85.00 percent and 84.92 percent,
respectively. Rahib et al. [7] conducted this study to investigate how Bangladeshis are reacting to and dealing
with the coronavirus disease (COVID-19) scenario. In this investigation, the status and comments on
COVID-19 concerns were gathered from multiple Facebook pages and YouTube channels run by reputable
Bangladeshi news organizations and health specialists. Throughout the study, a variety of machine learning
algorithms were studied, ranging from conventional algorithms like support vector machine and random
forest to deep learning algorithms like convolutional neural networks and long short-term memory.
Experiments were carried out on a 10,581-data-point categorized data set belonging to the authors. When
evaluating the performance of various models in terms of model assessment, the results demonstrate that long
short-term memory exceeds all of them, with an accuracy of 84.92 percent. To detect the polarity of textual
Facebook posts in Bangla containing people's points of view on Bangladesh Cricket, Faruque et al. [8]
proposed a sentiment polarity detection approach that uses three popular supervised machine learning
algorithms: naive Bayes (NB), support vector machines (SVM), and logistic regression (LR). With an
accuracy of 83 percent when considering n-gram as a feature, LR outperformed SVM and NB. Iqbal et al. [9]
proposed a four-step process for categorizing six emotions in Bengali literature, including data crawling, pre-
processing, labelling, and verification, with 7,000 texts labeled into six basic emotion groups. The dataset is
graded with a score of 0.969. Cohen's score reflects the close collaboration between corpus annotators and
experts. According to the analysis of appraisal, the distribution of emotion words also follows Zipf's law. The
BEmoC study's findings were also presented in terms of coding consistency, emotion density, and the most
utilized emotion words.
Shetu et al. [10] established a paradigm for parsing text data in paragraphs. To extract sentiment
from a text, they employed the bag of words method and lexical analysis method. Mamun et al. [11]
demonstrated that the ensemble approach (i.e., logistic regression+random forest+support vector machine)
with frequency-inverse document frequency (unigram+bi-gram+trigram) features outperformed the other
classifier models on the developed dataset, achieving the highest accuracy of 82 percent. Most of the
emotions conveyed on social media platforms are expressed through writing (such as status, tweets,
comments, and reviews). presents an ensemble-based method for categorizing Bengali textual sentiment into
positive and negative categories. Because the Bengali sentiment corpus was unavailable, this effort
additionally created a dataset called "Bengali sentiment analysis dataset". Neethu and Rajasree [12]
attempted to assess the sentiment of Twitter posts in a particular domain. They suggested a new feature
Sentiment analysis on Bangla conversation using machine learning approach (Mahmudul Hassan)
5564  ISSN: 2088-8708

vector that can differentiate between positive and negative sentiment in tweets. In order to examine twitter
data for sentiment analysis, Jain and Dandannavar [13] used naive Bayes and decision tree machine learning
methods. Because it is scalable and fast, their proposed model employs Apache Spark. Rahman and Dey [14]
provide two freely accessible Bangla datasets for sentiment analysis based on aspects. One dataset contains
user comments regarding cricket that have been human-annotated, while the other features restaurant
customer reviews. They also presented a fundamental method for analyzing our datasets utilizing the aspect
category extraction subtask.

3. RESEARCH METHOD
Research section will illustrate the overall architecture of our proposed system. The research method
is listed in Figure 1 as data collection, data pre-processing, model selection, statistical analysis, and its
implementation will be discussed in this portion. In Figure 1 the full method at a glance is shown.

Figure 1. Method at a glance

3.1. Data collection procedure

From various Bangla movies and short film scripts, we collected conversation data for our research
work. These conversations covered a large scale of topics like food, family, motivation, fraud, business, and
friends. After analyzing those collected data, we will split it into two categories: positive and negative. We
have collected about 1,141 data. These conversations include emotions like happy, sad, anger, worried, and
afraid. These categories help us to differentiate the whole dataset into two main categories of Positive and
Negative. Among 1,141 data there was 570 data for positive sentiment and for negative it was 571 data.
Figures 2 and 3 shows the sample dataset.

Figure 2. Sample data

Int J Elec & Comp Eng, Vol. 12, No. 5, October 2022: 5562-5572
Int J Elec & Comp Eng ISSN: 2088-8708  5565

Figure 3. Class label distribution

3.2. Data preprocessing and organizing

Firstly, we collect data from scripts and store them into an xlsx file. The dataset we have collected
has two attributes. These are positive and negative. As we already discussed, we collect data from movie and
short film scripts as conversation. Every conversation starts with a single word or single sentence. People can
express their feelings, emotions, and thoughts through a single word or sentence. To classify these
expressions into two main attributes we merged happiness, joy, motivation, and thankfulness into positive
conversations, and for negative conversation we merged sad, anger, backbiting, and worries. During
pre-processing, we remove punctuation in the first step. In natural language processing, for every language, it
is essential to identify and remove stop words. For our research work, we have collected Bangla stop words
and removed them to clean our data. There were about 410 stop words in the Bangla language. For example:
‘অতএব’, ‘অথচ’, ‘এই’, ‘একই’, ‘একটি’, ‘হয়’, ‘হয়ততো’, ‘ককন্তু’, ‘কী’, and ‘কক’. Here, Figure 4. shows the python
code for removing Bangla stop words and punctuations and Figure 5. shows the cleaned data what we
pre-processed.

Figure 4. Removing stop words and punctuations

Figure 5. Cleaned data

Sentiment analysis on Bangla conversation using machine learning approach (Mahmudul Hassan)
5566  ISSN: 2088-8708

To extract features from each of the conversations, several words and a number of characters are
needed. Figure 6 shows the result, respectively. After preprocessing procedure label encoding method applied
to the sentiment column. And then a pickle file generated. The pickle file contains temporary data for reuse
and also saves time during runtime execution. In this work, our cleaned data is stored as a pickle file for
upcoming procedures. We need to demonstrate our dataset data where highlights are age, occupation, house
type, want to switch jobs and we are giving low highlighting to other attributes. In Figure 7, cleaned data
along with counts of each conversation length and character is shown.

Figure 6. Word frequency and character frequency

Figure 7. Sample of cleaned dataset

Int J Elec & Comp Eng, Vol. 12, No. 5, October 2022: 5562-5572
Int J Elec & Comp Eng ISSN: 2088-8708  5567

3.3. Machine learning algorithms and statistical analysis

About 571 records for positive and 570 records are for negative conversations in our dataset. For the
dataset splitting purpose we used train-test split function. We followed supervised machine learning
techniques. To train our model we used 80% of our data and for test 20% of data used. In number, 912 data
used for trains and 229 data used for test purposes. To know the accuracy on our dataset we applied some
classifier-based algorithms. These are support vector machine, multinomial naïve Bayes, k-nearest neighbors,
logistic regression, decision tree, random forest, and stochastic gradient descent. In Figure 8, we have shown
that how we have done our research shortly details.

Figure 8. Proposed model structure

3.3.1. Feature extraction

We employ machine learning methods here to achieve natural language processing goals. Our model
is trained by extracting all characteristics of each phrase from two primary characteristics. A method called
tokenizer is presented here for this technique. Tokenizer divides phrases into words parts. These unique and
common words have identical properties. In addition, TF-IDF is also such a numerical figure that examines
the requirement of a term in a text. This approach is used by some important publications for several
languages. Their success inspired us, and we found that our learning algorithms were the most accurate.

3.3.2. Classifier algorithms

It builds numerous decision trees during training. The naïve Bayes classification presupposes that
there is no connection between the existence of a certain characteristic in a class and the presence of any
other characteristic. This model is straightforward to create and beneficial for very big datasets in particular.
Naïve Bayes even exceeds advanced categorization algorithms. The logical regression model may create a
probability model from a class or event. To decide, for example, one group of images including photographs
of different animals which may be investigated on a model of various classes. Stochastic gradient descent is
renowned for improving any method transmitted particularly in machine learning algorithms in order to
identify associated model parameters for both expected and actual results.

4. EXPERIMENTAL RESULT AND ANALYSIS

In this modern era, in intelligent analyzing of data and developing the related smart applications, the
understanding of IoT [15]–[17], cyber-security [18], in particular, machine learning and deep learning
[19]–[25] are crucial. According to our requirement, we update our model and dataset using machine learning
approach. From this modification, we can accomplish that our used classifier is exactly usable for a wide
range of use according to our dataset. As per our expectations, we achieved 86% accuracy from our proposed
mode which is a fruitful outcome. This performance of the model creates a path to think about the
improvement in results.
The research result was focused to identify whether a conversation is positive or negative. We have
applied classifiers based on different machine learning models to extract the conversation type. The result has
two criteria of positive and negative. There were 1141 data for training each of the models. We get various
Sentiment analysis on Bangla conversation using machine learning approach (Mahmudul Hassan)
5568  ISSN: 2088-8708

accuracy on different models. Among 7 models the support vector machine and multinomial naive Bayes
perform well with the highest accuracy. As we already discussed, we collect data from scripts as a
conversation. All conversations have people's emotions like happy, sad, worried, annoyed, and motivated.
We merged and categorized them into two main types, positive and negative. The decision-making capability
of the classifiers was measured by their performance. Accuracy, precision, recall, and F-score were used to
determine the performance of classifiers. For a classifier, the overall accuracy was considered an adequate
standard. In the test set, it is necessary to have a notion of the correctly classified samples.
In Table 1 the accuracy scores obtained for the classifiers built are given. Here it is clear that the
support vector machine gives the highest accuracy score of 0.85589 and multinomial naive Bayes gives
almost similar accuracy of 0.8513. That is why it was needed to calculate the other performance measures to
decide a suitable classifier for our dataset.
To measure the class agreement of the data labels with the positive labels given by the classifier the
precision is used. We have to calculate the precision scores for each of the two-class labels because it is
directly relevant to class labels. In Table 2 the values for each of the classifiers are given along with the 2
labels we used in this research work. We can see that the classifier random forest gives a score of 0.93 and
multinomial naive Bayes gives 0.85 for positive conversation.
To identify class labels recall is known as sensitivity of the measurement that represents the
effectiveness of the classifier. We also concentrated on achieving a score near 1 for the positive class label.
The recall scores for two-class labels and classifiers are reported in Table 3. The decision tree and support
vector machine had a recall score of 0.92 for positive dialogue. F1-score can be used to determine the
relationship between positive labels and those provided by the classifier. The harmonic means of precision
and recall for all two labels across all classifiers can be used to calculate it. The score close to 1 for the
positive class label was considered when determining the optimum model of classifier. Table 4 shows the F1
scores for the class labels. Vector machines and multinomial naïve classifiers are supported by the classifiers.
Bayes and stochastic gradient descent are the most effective methods for determining the best classifier for
our dataset.

Table 1. Accuracy of classifiers Table 2. Precision of classifiers

Classifier Accuracy Classifier Precision
random forest 74.24% Random forest 67.01%
decision tree 76.42% Decision tree 69.62%
logistic regression 82.53% Logistic regression 79.23%
k-nearest neighbors 82.97% K-nearest neighbors 79.39%
stochastic gradient descent 83.41% Stochastic gradient descent 79.55%
Multinomial naïve Bayes 85.15% Multinomial naïve Bayes 85.96%
Support vector machine 85.59% Support vector machine 81.68%

Table 3. Recall of classifiers Table 4. F1-score of classifiers

Classifier Recall Classifier F1-Score
Random forest 96.55% Random forest 79.15%
Decision tree 94.83% Decision tree 80.29%
Logistic regression 88.79% Logistic regression 83.74%
K-nearest neighbors 89.66% K-nearest neighbors 84.21%
Stochastic gradient descent 90.52% Stochastic gradient descent 84.68%
Multinomial naïve Bayes 84.48% Multinomial naïve Bayes 85.22%
support vector machine 92.24% Support vector machine 86.64%

Our objective is to predict the mentally hampered individuals with higher precision which was
achieved by random forest, multinomial naïve Bayes, and support vector machine. With remarkable accuracy
support vector machine, multinomial naïve Bayes, and stochastic gradient descent perform well among the
classifiers as shown in Table 5. Support vector machine, multinomial naive Bayes, and random forest all
perform well as individual classifiers, as seen in the tables. Support vector machines work well for the
challenge because our dataset is significantly more condensed, and the labels are poorly understood.
K-nearest neighbor works effectively since there are fewer dimensions or attributes. The assumption of class
conditional independence will only work for a large dataset, which is why the decision tree performs poorly
in this case.
To avoid over fitting and robustness, it is needed to have a strong correlation over fitting nuts,
though it is not exceptional. As it is not robust to noise and does not generalize well, future observed data
decision trees do not work too well. In Figure 9 the overall performance comparison is shown.

Int J Elec & Comp Eng, Vol. 12, No. 5, October 2022: 5562-5572
Int J Elec & Comp Eng ISSN: 2088-8708  5569

Table 5. Performance analysis of different algorithms

Classifier Accuracy Precision Recall F1-Score
Random forest 74.24% 67.01% 96.55% 79.15%
Decision tree 76.42% 69.62% 94.83% 80.29%
Logistic regression 82.53% 79.23% 88.79% 83.74%
k-nearest neighbors 82.97% 79.39% 89.66% 84.21%
Stochastic gradient descent 83.41% 79.55% 90.52% 84.68%
Multinomial naïve Bayes 85.15% 85.96% 84.48% 85.22%
Support vector machine 85.59% 81.68% 92.24% 86.64%

Figure 9. Performance analysis

4.1. Prediction
We have tried to test our model by using a random conversation data and we got a result.
In Figures 10 and 11, We can see positive and negative prediction conversation. That Mean’s, we can see that
our proposed model can extract sentiment from Bangla conversation data.

Figure 10. Predicting positive conversation

Figure 11. Predicting negative conversation

Sentiment analysis on Bangla conversation using machine learning approach (Mahmudul Hassan)
5570  ISSN: 2088-8708

5. CONCLUSION
This research work concludes with an expected outcome using machine learning approach of
extracting sentiment from Bangla conversation data. Text mining and text analysis are very new terms in
Bangla language. Though it is a tough task to work with some limitations, lacking the resources we tried to
overcome these difficulties. Technology makes the communication sector easier with advancement. But
embracing the advancement by ensuring the control of enormous data is necessary for us. We should be
concerned about these terminologies to make the world of data more accessible and convenient.

6. FUTURE WORK
This research work proposes a methodology that finds the scopes to work with Bangla conversation
data. To accomplish that, machine learning models were trained from Bangla conversation data and able to
extract sentiment from those conversations. There is a scope to apply a deep learning approach in our dataset
to improve efficiency. Here in this work, we extract sentiment as a positive and negative category. But on a
large scale, people’s emotions, and sentiments as individuals like sadness, anger, neutral, happiness, and fear
can also be extracted. For real-time conversation data, converting real-time conversations into text and
analyzing sentiment from these conversations can also be done. However, scope lies in every possible
opportunity. And opportunity revealed innovation and evolutions.

REFERENCES
[1] C. O. Alm, D. Roth, and R. Sproat, “Emotions from text,” in Proceedings of the conference on Human Language Technology and
Empirical Methods in Natural Language Processing, 2005, pp. 579–586, doi: 10.3115/1220575.1220648.
[2] C. Lin and Y. He, “Joint sentiment/topic model for sentiment analysis,” in Proceeding of the 18th ACM conference on
Information and knowledge management, 2009, 375, doi: 10.1145/1645953.1646003.
[3] T. Nasukawa and J. Yi, “Sentiment analysis: capturing favorability using natural language processing,” in Proceedings of the 2nd
International Conference on Knowledge Capture, K-CAP 2003, 2003, pp. 70–77, doi: 10.1145/945645.945658.
[4] A. L. Maas, R. E. Daly, P. T. Pham, D. Huang, A. Y. Ng, and C. Potts, “Learning word vectors for sentiment analysis,” in
Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, 2011,
vol. 1, pp. 142–150.
[5] N. R. Bhowmik, M. Arifuzzaman, and M. R. H. Mondal, “Sentiment analysis on Bangla text using extended lexicon dictionary
and deep learning algorithms,” Array, vol. 13, Mar. 2022, doi: 10.1016/j.array.2021.100123.
[6] T. T. Aurpa, R. Sadik, and M. S. Ahmed, “Abusive Bangla comments detection on Facebook using transformer-based deep
learning models,” Social Network Analysis and Mining, vol. 12, no. 1, Dec. 2022, doi: 10.1007/s13278-021-00852-x.
[7] M. R. H. K. Rahib, A. H. Tamim, M. Z. Tahmeed, and M. J. Hossain, “Emotion detection based on Bangladeshi people’s social
media response on COVID-19,” SN Computer Science, vol. 3, no. 2, Mar. 2022, doi: 10.1007/s42979-022-01077-1.
[8] M. A. Faruque, S. Rahman, P. Chakraborty, T. Choudhury, J.-S. Um, and T. P. Singh, “Ascertaining polarity of public opinions
on Bangladesh cricket using machine learning techniques,” Spatial Information Research, vol. 30, no. 1, pp. 1–8, Feb. 2022, doi:
10.1007/s41324-021-00403-8.
[9] M. D. A. Iqbal, A. Das, O. Sharif, M. M. Hoque, and I. H. Sarker, “BEmoC: a corpus for identifying emotion in Bengali texts,”
SN Computer Science, vol. 3, no. 2, Mar. 2022, doi: 10.1007/s42979-022-01028-w.
[10] S. F. Shetu, M. Saifuzzaman, M. Parvin, N. N. Moon, R. Yousuf, and S. Sultana, “Identifying the writing style of bangla language
using natural language processing,” in 2020 11th International Conference on Computing, Communication and Networking
Technologies (ICCCNT), Jul. 2020, pp. 1–6, doi: 10.1109/ICCCNT49239.2020.9225670.
[11] M. M. R. Mamun, O. Sharif, and M. M. Hoque, “Classification of textual sentiment using ensemble technique,” SN Computer
Science, vol. 3, no. 1, Jan. 2022, doi: 10.1007/s42979-021-00922-z.
[12] M. S. Neethu and R. Rajasree, “Sentiment analysis in Twitter using machine learning techniques,” in 2013 Fourth International
Conference on Computing, Communications and Networking Technologies (ICCCNT), 2013, pp. 1–5, doi:
10.1109/ICCCNT.2013.6726818.
[13] A. P. Jain and P. Dandannavar, “Application of machine learning techniques to sentiment analysis,” in 2016 2nd International
Conference on Applied and Theoretical Computing and Communication Technology (iCATccT), 2016, pp. 628–632, doi:
10.1109/ICATCCT.2016.7912076.
[14] M. Rahman and E. K. Dey, “Datasets for aspect-based sentiment analysis in bangla and its baseline evaluation,” Data, vol. 3,
no. 2, May 2018, doi: 10.3390/data3020015.
[15] M. Saifuzzaman, S. F. Shetu, N. N. Moon, F. N. Nur, and M. H. Ali, “IoT based street lighting using dual axis solar tracker and
effective traffic management system using deep learning: bangladesh context,” in 2020 11th International Conference on
Computing, Communication and Networking Technologies (ICCCNT), Jul. 2020, pp. 1–5, doi:
10.1109/ICCCNT49239.2020.9225590.
[16] M. Saifuzzaman, N. N. Moon, and F. N. Nur, “IoT based street lighting and traffic management system,” in 2017 IEEE Region 10
Humanitarian Technology Conference (R10-HTC), Dec. 2017, pp. 121–124, doi: 10.1109/R10-HTC.2017.8288921.
[17] R. Hasan, S. Islam, M. H. Rahman, M. Saifuzzaman, S. F. Shetu, and N. N. Moon, “Implementation of low cost real-time
attendance management system: a comparative study,” in 2020 8th International Conference on Reliability, Infocom Technologies
and Optimization (Trends and Future Directions) (ICRITO), Jun. 2020, pp. 1098–1101, doi:
10.1109/ICRITO48877.2020.9197764.
[18] S. F. Shetu, M. Saifuzzaman, N. N. Moon, and F. N. Nur, “A survey of botnet in cyber security,” in 2019 2nd International
Conference on Intelligent Communication and Computational Techniques (ICCT), Sep. 2019, pp. 174–177., doi:
10.1109/ICCT46177.2019.8969048.
[19] K. K. Podder et al., “Bangla sign language (BdSL) alphabets and numerals classification using a deep learning model,” Sensors,

Int J Elec & Comp Eng, Vol. 12, No. 5, October 2022: 5562-5572
Int J Elec & Comp Eng ISSN: 2088-8708  5571

vol. 22, no. 2, Jan. 2022, doi: 10.3390/s22020574.

[20] M. Hossain et al., “Prediction on domestic violence in Bangladesh during the COVID-19 outbreak using machine learning
methods,” Applied System Innovation, vol. 4, no. 4, Oct. 2021, doi: 10.3390/asi4040077.
[21] M. Al-Smadi, O. Qawasmeh, B. Talafha, and M. Quwaider, “Human annotated Arabic dataset of book reviews for aspect based
sentiment analysis,” in 2015 3rd International Conference on Future Internet of Things and Cloud, Aug. 2015, pp. 726–730, doi:
10.1109/FiCloud.2015.62.
[22] L. Khan, A. Amjad, K. M. Afaq, and H.-T. Chang, “Deep sentiment analysis using CNN-LSTM architecture of English and
Roman Urdu text shared in social media,” Applied Sciences, vol. 12, no. 5, Mar. 2022, doi: 10.3390/app12052694.
[23] M. Chen, K. Ubul, X. Xu, A. Aysa, and M. Muhammat, “Connecting text classification with image classification: a new
preprocessing method for implicit sentiment text classification,” Sensors, vol. 22, no. 5, Feb. 2022, doi: 10.3390/s22051899.
[24] K. Schouten, O. van der Weijde, F. Frasincar, and R. Dekker, “Supervised and unsupervised aspect category detection for
sentiment analysis with co-occurrence data,” IEEE Transactions on Cybernetics, vol. 48, no. 4, pp. 1263–1275, Apr. 2018, doi:
10.1109/TCYB.2017.2688801.
[25] H. Zou and K. Xiang, “Sentiment classification method based on blending of emoticons and short texts,” Entropy, vol. 24, no. 3,
Mar. 2022, doi: 10.3390/e24030398.

BIOGRAPHIES OF AUTHORS

Mahmudul Hassan studied Computer Science and Engineering from

Daffodil International University. His main interests in research fields are natural
language processing, image processing, machine learning and data mining. His research
interests also include data science and computer vision. He can be contacted at email:
[email protected].

Shahriar Shakil studied Computer Science and Engineering from Daffodil

International University. His main interests in research fields are image processing,
machine learning and data mining. His research interests also include data science and
computer vision. He can be contacted at email: [email protected].

Nazmun Nessa Moon is an associate professor of the Department of

Computer Science and Engineering at Daffodil International University. She received the
B.Sc. degree in Computer Science and Engineering from Rajshahi University of
Engineering and Technology and M.Sc. in Information and Communication Technology
from Bangladesh University of Engineering and Technology (BUET). Her interested
research fields are IoT, digital image processing and machine learning. She can be
contacted at email: [email protected].

Mohammad Monirul Islam is a lecturer (senior scale) of the Department

of Computer Science and Engineering at Daffodil International University. He completed
his B.Sc. (India) in Computer Science and M.Sc. (UK) in Computing. His areas of
interests are navigation for organizations/institutions and data warehousing. He can be
contacted at email: [email protected].

Sentiment analysis on Bangla conversation using machine learning approach (Mahmudul Hassan)
5572  ISSN: 2088-8708

Refath Ara Hossain is a lecturer of the Department of Computer Science

and Engineering at Daffodil International University. Her interested research fields are
data mining, digital image processing and machine learning. She can be contacted at
email: [email protected].

Asma Mariam is a lecturer of the Department of Computer Science and

Engineering at Daffodil International University. Her interested research fields are digital
image processing, data mining and machine learning. She can be contacted at email:
[email protected].

Fernaz Narin Nur is an associate professor at the Notre Dame University

Bangladesh (NDUB) in the Department of CSE. She is a passionate researcher in the
fields of wireless sensor network, cloud computing, internet of things, and performance
analysis. She can be contacted at email: [email protected].

Int J Elec & Comp Eng, Vol. 12, No. 5, October 2022: 5562-5572