SENTIMENT ANALYSIS OF BANGLA-ENGLISH CODE-MIXED AND TRANSLITERATED SOCIAL MEDIA COMMENTS USING MACHINE LEARNING

Tianjin Daxue Xuebao (Ziran Kexue yu Gongcheng Jishu Ban)/
Journal of Tianjin University Science and Technology
ISSN (Online):0493-2137
E-Publication: Online Open Access
Vol: 56 Issue: 11:2023
DOI: 10.5281/zenodo.10152990
Nov 2023 | 26
SENTIMENT ANALYSIS OF BANGLA-ENGLISH CODE-MIXED AND
TRANSLITERATED SOCIAL MEDIA COMMENTS USING MACHINE
LEARNING
FAHIMA HOSSAIN
Lecturer, Department of Computer Science & Engineering, Hamdard University Bangladesh, Bangladesh.
E-mail: minda.fahima25@gmail.com
NUSRAT JAHAN
BSc in CSE, Hamdard University Bangladesh, Bangladesh. E-mail: nusrat.jahan.az@gmail.com
JOYREYA ABADIN
BSc in CSE, Hamdard University Bangladesh, Bangladesh. E-mail: ajoyreya285@gmail.com
Abstract
In this era of technology, online communication has expanded to incorporate different language
combinations and expressive styles. Code-mixing, or combining languages such as Bengali and English,
is frequent in multilingual environments. In addition, the use of emojis, which are small graphical icons, has
increased the complexity of online communication by allowing for more emotional expression. Sentiment
analysis of comments on social media that are combined with emojis and written in Bangla-English
language is the main topic of this work. This research aims to understand and classify the sentiments
expressed in code-mixed comments, including the nuanced role of emojis. The data was preprocessed
first, then feature extraction was performed using the TF-IDF Vectorizer and CountVectorizer algorithms.
For the analysis, nine different machine learning algorithms were used. With a remarkable accuracy of
85.7% and an F1 score of 85.0%, the Support Vector Classifier stood out as the most successful model.
This highlights the utility of including emoji-based features for complex sentiment analysis, particularly when
dealing with code-mixed data. The dataset contained 2055 comments from Facebook pages, including
Bangla-English comments with and without emojis and comments that only contained emojis. To prepare
the dataset for analysis, preprocessing methods included removing irrelevant data and converting emojis
into Unicode short names.
Index Terms: Sentiment Analysis, Emoji Analysis, Natural Language Processing, Machine Learning, Code-
mixing, Emoji, Unicode.
1 INTRODUCTION
Today, individuals effortlessly blend languages while communicating online, which is
referred to as code-mixing, in order to express their thoughts and emotions. People who
are fluent in multiple languages often blend them together, incorporating English elements
into their native language. The fusion of English and Bengali, known as Benglish, is a
prime example of this phenomenon. It enables individuals to convey their thoughts and
ideas by using a blend of both languages. This linguistic phenomenon highlights the ever-
changing nature of communication within multilingual groups [1].
Sentiment analysis, a crucial area of natural language processing (NLP), focuses on
understanding and classifying the emotions and points of view expressed in textual data.
According to the analytic context, these attitudes are divided into a variety of groups,

Vol: 56 Issue: 11:2023
DOI: 10.5281/zenodo.10152990
Nov 2023 | 27
including negative, neutral, and positive, with the possibility to further segment them into
extremely negative, strongly positive, and other categories [2]. This technique has a wide
range of applications, from business insights to election forecasts, by decoding feelings
in customer feedback, social media posts, and public opinion [3, 4]. However, whereas
sentiment analysis research thrives in English, Bangla lags behind due to linguistic
difficulty and a lack of resources [5].
Emojis have revolutionized modern communication and have become a common feature
in social media conversations [6]. Through the effective communication of emotions,
context, and intent, these Unicode symbols cross linguistic and cultural boundaries [7, 8,
9]. On Messenger, more than 900 million emoticons are exchanged every day, while
Facebook sees more than 700 million emojis shared in a day [10]. Emojis can substitute
for words and represent a variety of ideas, including emotions, objects, activities, and
more because of their adaptability [11]. Emoji usage is widespread, but little study has
been done on it [12, 13, 14]. Emojis and text are combined in this study, which recognizes
the variety of expressions they provide. Including emojis in textual sentiment analysis is
beneficial. They provide a universal language [7, 8, 9] for expressing emotions, enhance
prediction accuracy, and align with contemporary communication styles.
This study collects Facebook data, which consists of Bangla-English comments with and
without emojis, along with comments that only contain emojis. The dataset contains a
total of 2055 labeled points and has been preprocessed and balanced using TF-IDF and
Count Vectorizer before being split into training and testing. To perform sentiment
analysis, nine categorization algorithms are utilized.
This work aims to fill the research gap on extracting sentiment from code-mixed Bengali-
English texts with emojis in social media. The contributions of this study are as follows:
1. Created a dataset of tagged Bangla-English code-mix with added emojis.
2. Managed imbalanced data properly to reduce overfitting.
3. Created a customized Bangla-English lemmatizer.
4. Developed an empirical model for sentiment analysis.
This research is divided into numerous Sections that methodically cover various parts of
the investigation. Section 2 offers a concise summary of related works to determine the
current state of knowledge in the subject matter. In Section 3, a summary of the data-
gathering process is provided, including details on data storage and the methodology
used for the study. Furthermore, Section 4 evaluates the proposed model and presents
its outcomes, providing significant information on its effectiveness. The findings of the
paper are presented in Section 5, which also suggests possible directions for further
research.

Vol: 56 Issue: 11:2023
DOI: 10.5281/zenodo.10152990
Nov 2023 | 28
2. LITERATURE REVIEW
Mahmud AA et al. [15], proposed an approach to sentiment analysis that supports Bangla-
English code-mixing and transliterated features. The authors pre-processed the data by
stopword remove, punctuation remove, pos tagging and for word embedding they used
glove, and word2vector, and after all that, LSTM and BERT were applied.
Khan et al. [16], proposed sentiment analysis approach for Bangla language and worked
with five different emotions: Happy, Sad, Angry, Surprised, and Excited. Additionally, their
paper also deals with two categories: Abusive and Religious. The authors used the TFiDF
method for converting the data into weighted vector form. The total work has done in two
experimental steps. In the first experiment, they trained the dataset with algorithms. In the
second experiment, they mapped their 7 classes (Sad, Happy, Angry, Excited, Religious,
Abusive, Surprised) into three higher classes (Positive, Negative, and Neutral). The best
accuracy (62%) for 7 class achieved using Support Vector Machine (SVM). And for 3
classes, 73% is the best accuracy obtained using Support Vector Machine (SVM).
Khan et al [17], proposed a model for depression analysis from Bangla post via Natural
Language Processing (NLP). For data cleaning they added contraction, removed regular
expression, removed stop words and tokenized the data using CountVectorizer. The
Multinomial Naive Bayes algorithm provide best accuracy for their model.
Al-Azani et al. [2], introduced a approach by incorporating nonverbal features, specifically
emojis, for sentiment analysis of Arabic microblogs where each instances, contained at
least one emoji. For feature selection, they used ReliefF and Correlation-Attribution
Evaluator (CAE) as their approach was text-independent and would only focus on emojis
features. The study achieves an F1 score of 80.30% and an AUC (Area Under the Curve)
of 87.30% by selecting multinomial naive Bayes classifier and 250 of the most relevant
emojis. The authors indicate that using emoji-based features alone can be highly effective
in detecting sentiment polarity.
Mandal et al. [18], addressed sentiment analysis for code-mixed social media content.
They created a Bengali-English dataset and utilized hybrid approaches for language and
sentiment tagging, obtaining 81% accuracy in language identification and 80.97%
accuracy in sentiment classification. The research paper provides valuable resources for
code-mixed sentiment analysis, which could potentially inspire more research in this
complex field of natural language processing.
Singh et al. [1], describe their findings on sentiment analysis of Hinglish (English-Hindi
mix) code-mixed social media tweets. Data consolidation, cleansing, transformation, and
modeling are all part of the research. There were several vectorization approaches and
machine learning algorithms utilized. The SemEval-2020 dataset was divided into
positive, neutral, and negative feelings. The ensemble voting classifier attained the best
F1-score (69.07), suggesting the possibility for additional improvements employing neural
networks and advanced data processing techniques. The research underlines the need
to normalize Hindi spellings and discusses potential improvements for Hindi words using
neural networks and methods such as lemmatization and part-of-speech tagging.

Vol: 56 Issue: 11:2023
DOI: 10.5281/zenodo.10152990
Nov 2023 | 29
Alzubaidi et al. [14], employed heterogeneous data mining (HDM) for integrated
processing to explore sentiment analysis of tweets utilizing both text and emojis. The
results indicate that using HDM along with text or emojis helps improve sentiment analysis
accuracy. With Support Vector Machines (SVM), the accuracy reached 85.89%. This
highlights the significance of incorporating emojis to enhance bi-sense sentiment analysis
accuracy.
Tareq et al. [19], in this study the author's utilized a Bangla-English code-mixed (BE-CM)
dataset they created to address challenges related to sentiment analysis in low-resource
languages like Bangla. Their data augmentation technique improves cross-lingual
comprehension in code-mixed expressions, achieving an 87% weighted F1 score using
XGBoost and FastText embeddings. The work tackles the issues of sentiment analysis in
multilingual communities, particularly for languages with limited resources like Bangla.
Jamatia et al. [20], the paper research sentiment analysis in Indian language code-mixed
social media texts. For sentiment prediction in English-Hindi and English-Bengali, they
compare standard machine learning approaches with deep learning models such as
BiLSTM-CNN, Double BiLSTM, Attention-based, and BERT. The deep learning models
consistently outperformed the classical methods, which is pretty impressive. The
Attention-based model achieved the highest F1 scores of 67.5 (English-Bengali) and 60.4
(English-Hindi). It's interesting to note that while domain-specific BERT embeddings show
potential, pre-trained BERT embeddings are still pretty competitive. Overall, the study
highlighted the challenges in code-mixed sentiment analysis and emphasized the value
of deep learning as well as the need for more extensive datasets.
Limitations: The existing literature reveals a variety of approaches to sentiment analysis,
with some focusing solely on code-mix languages, others on emojis exclusively, and
some on text and emojis in native languages. However, there is still a lack of research
on Bangla-English code-mixed language, where English elements are integrated into
native language text through phonetic typing. Our study presents a sentiment analysis
framework for Bangla-English code-mixed language with emojis, as no previous
exploration has been conducted on this topic yet. The studies discussed in this section
are summarized in Table 1 based on various attributes.
Table 1: Overview of Relevant Studies Referenced from [1, 2 14-20].
Authors Contributions Datasets Class
Performances
(Best Score)
Mahmud
AA et al.
[15]
1. Applied two types of feature
extraction
methods to convert text into
vectors.
2. Plotted 100 dimensions’ vectors
into 2 dimensional graphs.
3. Used test set differently to
measure the performance of the
model. The first test set contains
10% code mix data and the second
Bkash app
review from
google playstore
(0.2 million
sample with 12%
bangla-english
mixed sentence)
Positivie,
Negative
Accuracy: 94%

Vol: 56 Issue: 11:2023
DOI: 10.5281/zenodo.10152990
Nov 2023 | 30
test set contains 42% code mix
data.
Khan et
al. [16]
1. Created dataset of pure Bangla
text or Bangla text mixed with
some characters.
2. Applied Tf-idf as feature
extraction
Facebook post
comments
Positivie,
Negative,
Neutral
Accuracy: 73%
Khan et al
[17]
1. Build a model to detect
depression related post.
2. Applied Tf-idf as feature
extraction
Bengali post of
facebook
Happy, Sad Accuracy:
86.67%
Al-Azani
et al [2]
1. Conducted research on Arabic
sentiment analysis to evaluate
the impact of different emojis
on predicting sentiment
polarity.
Arabic
microblogs
(2091 sample)
Positivie,
Negative
Accuracy:
80.34%
Soumil
Mandal et
al. [18]
1. Create a gold standard Bengali-
English code-mixed dataset.
2. Develop an accurate language
identification system with an
81% accuracy rate.
3. The dataset, which includes
code-mixed text and is
annotated with language and
polarity tags, is now accessible
in JSON format for future
research purposes.
Gold standard
Bengali-English
code-mixed
dataset
(600 tweets)
Positive,
Negative,
And
Neutral
Accuracy:
80.97%
Singh et
al. [1]
1.Applied different data-
cleaning approaches,
transformation techniques, and
machine-learning algorithms.
2. Identified areas for growth and
development, such as neural
network exploration and spell
normalization.
Task 9:
Sentiment
Analysis of
Code-Mixed
Social Media
Text (Hinglish)
(14,000 tweets)
[21]
Positive,
Negative,
And
Neutral
F1-score:
69.07%
Alzubaidi
et al. [14]
1. Uses texts and emojis to
demonstrate HDM's
supremacy for sentiment
analysis.
2. Highlights the efficacy of emojis
even in small quantities.
3. Introduces a uniform approach
for both text and emoji
processing.
English corpus
composed of
16,207 tweets
[22]
Positive
and
Negative
Accuracy:
85.89%
Tareq et
al. [19]
1. Creating the BE-CM in code-
mixed Bangla-English text.
2. Evaluating several models and
BERT iterations for sentiment
categorization.
3. Introducing an effective data
augmentation strategy for
Gold standard
Bangla-English
code mix (BE-
CM) dataset
(18,074
Sentences)
Positive
state,
Negative
state,
Mixed
positive,
Mixed
F1- score: 87%

Vol: 56 Issue: 11:2023
DOI: 10.5281/zenodo.10152990
Nov 2023 | 31
cross-lingual sentiment
analysis.
negative
and Neutral
Jamatia et
al. [20]
1. Demonstrates the effectiveness
of deep learning in code-mixed
sentiment analysis.
2. Increse sentiment analysis in
multilingual contexts
ICON-2017 SAIL
Tool-Contest
Annotated
Corpora [43],
Joshi et al. [24],
SemEval 2013
Task 2B Twitter
Dataset [25],
Barnes et al. [26]
Positive,
Negative,
And
Neutral
F1-scores:
67.5%
(English-
Bengali) and
60.4%
(English-Hindi)
3. METHODOLOGY
The proposed model working mechanism to detect sentiment is present in this section.
Figure 1 illustrates the approach used in this model.
Fig. 1: Flow Diagram of the Proposed Model.
3.1 Data collection and dataset preparation
Machine learning heavily relies on data as a fundamental component. The data used to
utilize this paper is the comments of general people that have been collected manually
from different Facebook pages using facepager. Facepager uses a JSON-based API to
fetch publicly accessible data from Facebook, YouTube, and Twitter. An access token is
required, generated automatically when a profile is logged in. An additional method for
obtaining access tokens is through the utilization of a website called the 'Graph API

Vol: 56 Issue: 11:2023
DOI: 10.5281/zenodo.10152990
Nov 2023 | 32
Explorer - Facebook Developer.' Facepager exports raw data as an Excel CSV sheet and
stores it in a SQLite database (2).
2055 comments have been collected that are mainly Bangla, written using English letters
(generally known as Banglish), mix of Banglish comments with or without emoji, or only
emoji. Along with these, some commonly used English words and sentences that are
mostly used (such as: ok, well done, good, problem, best of luck, etc) are also consumed
as collected comment. The collected comments are then exported as an Excel CSV
sheet. Each comment has been labeled manually based on its real aspect of use. Overall
nine columns are present in the dataset that is explained in Table 2. Figure 1 represents
the sample scenario of the dataset.
Table 2: Columns included in the dataset.
Column Description
Page_link The page link of particular post from which comments are collected
Page_id The page id of particular post from which comments are collected
Query_time The time when the comments have collected
Post_id The post id of post from which comments are collected
Post_created_time The post created time of post from which comments are collected
Post The post from which comments are collected
Comment_created_time Commented time
Comment Comment of general people
Sentiment Class (positive, Negative, Neutral) of comment
Fig. 2: The sample scenario of the dataset.

Vol: 56 Issue: 11:2023
DOI: 10.5281/zenodo.10152990
Nov 2023 | 33
3.2 Data preprocessing
To enhance the performance of the machine learning model raw data needs to be
preprocessed. At the very first step of data preprocessing, keeping only comments and
sentiment columns, all other unnecessary columns have been dropped. The
preprocessing steps followed for this model are:
1. Handling missing data: The handling of missing data operations is carried out on the
entire dataset. As missing data can bias the result of the model, If a row with a missing
value is detected, the entire row is dropped.
2. Remove non relevant information: hyperlink, URL, email, name mention, hashtag,
number, new lines, and extra space have no significant rules for detecting sentiment.
So, these are removed from the comments.
3. Convert in lowercase: To avoid any case sensitivity all letters are converted into
lowercase.
4. Remove punctuations: Punctuation as like ('!"#$%&'()*+,-./:;<=>?@[]^_`{|}~') has
been removed from comments.
5. Convert emoji to Unicode CLDR: Each emoji has a unicode CLDR short name, like
(😢 | crying face). For better performance in the model, emojis are converted into
Unicde CLDR shortname.
6. Remove duplicate values: This operation is done in two steps. In the first step, if any
word contains sequentially repeated letters more than twice, then the repeated letters
have been removed. Such as (onnnnkkk-onk). In the second step, if any word is
sequentially repeated, then the repeated word has been removed. Such as (sokal
sokal egula valo lage na - sokal egula valo lage na).
3.3 Tokenization
After completing the data preprocessing steps to convert the raw dataset into a clean
dataset, each data is converted into a list of tokens.
3.4 Normalization
In Bangla-English code mixed sentences there are various forms of words that are spelled
differently but pronounced similarly which can be converted to one single form. As like
(valo-vlo-bhalo, kharap-khrp etc). Another main things among words are, some people
like to write short form of that word and some write the enlarge word. For the proposed
model, normalization list have created manually with the words present in the dataset.
The first word was used for replacing the variations of the spellings of the same Banglish
word.

Vol: 56 Issue: 11:2023
DOI: 10.5281/zenodo.10152990
Nov 2023 | 34
Fig. 3: Sample of Data Preprocessing
Fig. 4: Sample of Created Normalization.
3.5 Feature Extraction
Each word are then transformed into numerical value using two different feature
extraction method to evaluate which performs better for the model. Using TF-IDF and
countVectorizer each unique token gets a feature index. Finally, each comment becomes
a vector, and the weighted numbers of each vector represent the score of features.

Vol: 56 Issue: 11:2023
DOI: 10.5281/zenodo.10152990
Nov 2023 | 35
3.6 Data Balancing
The dataset contains 887 Neutral, 559 Negative, and 609 Positive comments. This
implies that the classes of the dataset are imbalanced which can effects the performance
of the machine-learning model. To address the imbalanced data, there are several
different approaches that can be taken, including both undersampling and oversampling
techniques. Synthetic Minority Over-sampling Technique (SMOTE) has been used for this
model to make a balance class dataset.
Fig. 5: Class distribution of the dataset.
3.7 Dataset Distribution and Classification
From the dataset 80% kept for train the model and 20% kept for evaluate the performance
of model. For classifying comments on the basis of their sentiment into three classes:
Positive, Negative and Neutral, nine machine learning algorithm has used; Bernoulli naive
Bayes (BNB), Multinomial naive Bayes (MNB), Gaussian naive Bayes (GNB), Decision
Tree Classifier (DT), K-Nearest Neighbour (KNN), Logistic regression (LR), Linear
Support Vector Classifier (SVC), Support Vector Classifier (SVC) and random forests
(RF).
4. RESULT ANALYSIS
The accuracy of several algorithm used to train the model dataset is shown in Figure 6.
The barchart highlights that the algorithms obtained comparatively better accuracy using
Term Frequency and Inverse Document Frequency (TF-IDF) rather than CountVectorizer
feature extraction method. The highest accuracy, 85.7% obtained by Support Vector
Classifier (SVC) using Term Frequency and Inverse Document Frequency (TF-IDF). The
lowest accuracy obtained by K-Nearest Neighbour algorithm for both TF-IDF and
CountVectorizer.

Vol: 56 Issue: 11:2023
DOI: 10.5281/zenodo.10152990
Nov 2023 | 36
Fig. 6: Predictive accuracy barchart for TF-IDF and CountVectorizer.
For more depth analysis, three distinct metrics—Precision, Recall and F1-Score have
been identified for implemented classification algorithms to evaluate and better
comprehend the model's performance and predictive outcomes.
Fig. 7: Predictive Precision barchart for TF-IDF and CountVectorizer.

Vol: 56 Issue: 11:2023
DOI: 10.5281/zenodo.10152990
Nov 2023 | 37
Fig. 8: Predictive recall barchart for TF-IDF and CountVectorizer.
Table 3: Precision, Recall and F1-Score of algorithms using Countvectorizer.
Classification Precision Recall F1-Score
BernoulliNB 0.66 0.64 0.63
GaussianNB 0.55 0.52 0.49
MultinomialNB 0.71 0.70 0.69
DecisionTreeClassifier 0.68 0.68 0.66
K-Nearest Neighbors 0.60 0.57 0.52
LogisticRegression 0.74 0.74 0.73
Linear Support Vector Classifier 0.72 0.72 0.71
Support Vector Classifier 0.70 0.69 0.69
RandomForestClassifier 0.70 0.70 0.69
Table 3 and Table 4 shows the results for the nine considered classifiers using two
different feature extraction method.
Table 4: Precision, Recall and F1-Score of algorithms using TF-IDF.
Classification Precision Recall F1-Score
BernoulliNB 0.74 0.74 0.73
GaussianNB 0.70 0.67 0.64
MultinomialNB 0.79 0.79 0.78
DecisionTreeClassifier 0.72 0.72 0.71
K-Nearest Neighbors 0.76 0.58 0.52
LogisticRegression 0.81 0.81 0.81
Linear Support Vector Classifier 0.81 0.82 0.81
Support Vector Classifier 0.87 0.86 0.85
RandomForestClassifier 0.78 0.79 0.79

Vol: 56 Issue: 11:2023
DOI: 10.5281/zenodo.10152990
Nov 2023 | 38
5. CONCLUSION & FUTURE WORK
This study presents a new technique for sentiment analysis of Bangla-English code-mixed
language, including emojis, in social media communication. The study emphasizes the
challenges of sentiment analysis in code-mixed language. It also highlights the
importance of emojis for nuanced emotional interpretation, while acknowledging limited
previous research. This study contributes to sentiment analysis by constructing a dataset,
handling imbalanced data, creating a Bangla-English lemmatizer, and evaluating
classification algorithms extensively. The results show that using the Term Frequency
and Inverse Document Frequency (TF-IDF) feature extraction method with the Support
Vector Classifier (SVC) achieved an accuracy of 85.7% and an F1 score of 85.0%.
This study proposes several potential directions for future exploration in sentiment
analysis. These include expanding datasets, incorporating additional linguistic features,
improving preprocessing techniques, and optimizing language variations to enhance
effectiveness. Additionally, it aims to develop emotion recognition models that incorporate
emojis, stickers, GIFs, and other elements to examine emotions in code-mixed text.
These future research paths have the potential to not only improve the accuracy of
sentiment analysis but also deepen our understanding of sentiment dynamics in code-
mixing communities.
Acknowledgement
This project is financially supported by a research grant from Hamdard University Bangladesh.
References
1) Gaurav Singh, “Sentiment Analysis of Code-Mixed Social Media Text (Hinglish)”, School of
Computing, University of Leeds, Leeds, LS29JT, UK, doi.org/10.48550/arXiv.2102.12149.
2) Al-Azani S, El-Alfy ESM. Emoji-Based Sentiment Analysis of Arabic Microblogs Using Machine
Learning. In: 2018 21st Saudi Computer Society National Computer Conference (NCC) [Internet].
Riyadh: IEEE; 2018 [cited 2023 Aug 18]. p. 1–6. Available from:
https://ieeexplore.ieee.org/document/8592970/
3) GS. Solakidis et al. 2014. Multilingual sentiment analysis using emoticons and keywords. Proc. - 2014
IEEE/WIC/ACM Int. Jt. Conf. Web Intell. Intell. Agent Technol. - Work. WI-IAT 2014 2, (2014), 102–
109.
4) AJ Shamal. 2019. Sentiment Analysis using Token2Vec and LSTMs: User Review Analyzing Module.
(2019), 48–53.
5) Nusrath Tabassum and Muhammad Ibrahim Khan, ”Design an Empirical Framework for Sentiment
Analysis from Bangla Text using Machine Learning”, 2019 International Conference on Electrical,
Computer and Communication Engineering (ECCE), 7-9 February, 2019,
doi.org/10.1109/ECACE.2019.8679347
6) L. K. Kaye, S. A. Malone, and H. J. Wall, “Emojis: Insights, affordances, and possibilities for
psychological science,” Trends in cognitive sciences, vol. 21, no. 2, pp. 66–68, 2017
7) “Emoji: The 21st Century Language", Divyansh Kandpal FEB 14, 2021, 18:57 IST,
www.timesofindia.indiatimes.com/readersblog/dkwrites.

Vol: 56 Issue: 11:2023
DOI: 10.5281/zenodo.10152990
Nov 2023 | 39
8) “World Emoji Day 2023: Celebrating the universal language of emojis and their digital impact”, Govind
Choudhary, 17 Jul 2023, www.livemint.com/authors/govind-choudhary
9) “Emoji is the new universal language. And it’s making us better communicators”, Vyvyan Evans,
Professor of Linguistics | Researcher, published, Aug 5, 2017, Linkedin,
www.linkedin.com/pulse/emoji-new-universal-language-its-making-us-better-vyv-evans
10) “Emojipedia Shares Most Used Emojis of 2023 for World Emoji Day”, Andrew Hutchinson, Published
July 17, 2023, www.socialmediatoday.com/news/emojipedia-shares-most-used-emojis-2023-world-
emoji-day/688143/
11) P. K. Novak, J. Smailovic, B. Sluban, and I. Mozeti ´ c, “Sentiment of ˇ emojis,” PloS one, vol. 10, no.
12, p. e0144296, 2015.
12) https://en.oxforddictionaries.com/word-of-the-year/word-of-the-year-2015
13) G. Abdulsattar et al. 2018. Stock Market Classification Model Using Sentiment Analysis on Twitter
Based on Hybrid Naive Bayes Classifiers.
14) Mohammed Alzubaidi and Farid Bourennani. 2021. Sentiment Analysis of Tweets Using Emojis and
Texts. In 2021 The 4th International Conference on Information Science and Systems (ICISS 2021),
March 17–19, 2021, Edinburgh, United Kingdom. ACM, New York, NY, USA, 4 pages.
https://doi.org/10.1145/ 3459955.3460608
15) Mahmud AA. Artificial Intelligence in Business Decision Making: A Study on Code-Mixed and
Transliterated Bangla Customer Reviews. SSRN Electron J [Internet]. 2021 [cited 2023 Aug 25];
Available from: https://www.ssrn.com/abstract=3875534
16) Khan MdSS, Rafa SR, Abir AEH, Das AK. Sentiment Analysis on Bengali Facebook Comments To
Predict Fan’s Emotions Towards a Celebrity. J Eng Adv. 2021 Jul 23;118–24.
17) Khan MdRH, Afroz US, Masum AKM, Abujar S, Hossain SA. Sentiment Analysis from Bengali
Depression Dataset using Machine Learning. In: 2020 11th International Conference on Computing,
Communication and Networking Technologies (ICCCNT) [Internet]. Kharagpur, India: IEEE; 2020
[cited 2023 Aug 18]. p. 1–5. Available from: https://ieeexplore.ieee.org/document/9225511/
18) Soumil Mandal, Sainik Kumar Mahata and Dipankar Das, “Preparing Bengali-English Code-Mixed
Corpus for Sentiment Analysis of Indian Languages”, 11 Mar 2018,
https://doi.org/10.48550/arXiv.1803.04000
19) M. Tareq, M. F. Islam, S. Deb, S. Rahman and A. A. Mahmud, "Data-Augmentation for Bangla-English
Code-Mixed Sentiment Analysis: Enhancing Cross Linguistic Contextual Understanding," in IEEE
Access, vol. 11, pp. 51657-51671, 2023, doi: 10.1109/ACCESS.2023.3277787.
20) Jamatia, S. D. Swamy, B. Gambäck, A. Das, S. Debbarma .(2020). Deep Learning Based Sentiment
Analysis in a Code-Mixed English-Hindi and English-Bengali Social Media Corpus. International
Journal on Artificial Intelligence Tools. doi:10.1142/s0218213020500141.
21) Patwa, P., Aguilar, G., Kar, S., Pandey, S., Pykl, S., Garrette, D., Gambck, B., Chakraborty, T., Solorio,
T. and Das, A. 2020. SemEval-2020 Sentimix Task 9: Overview of SENTIment Analysis of Code-
MIXed Tweets. Proceedings of the 14th International Workshop on Semantic Evaluation (SemEval-
2020)
22) A.M. MacEachren et al. 2010. Geo-Twitter Analytics: Applications in Crisis Management. Proc. 25th
Int. Cartogr. Conf. (2010), 1–8.
23) B. G. Patra, D. Das, and A. Das. Sentiment analysis of code-mixed indian languages: An overview of
sail code-mixed shared task @icon-2017. CoRR, abs/1803.06745, 2018.

Vol: 56 Issue: 11:2023
DOI: 10.5281/zenodo.10152990
Nov 2023 | 40
24) Joshi, A. Prabhu, M. Shrivastava, and V. Varma. Towards sub-word level compositions for sentiment
analysis of Hindi-English code mixed text. In Proceedings of the 26th International Conference on
Computational Linguistics, pages 2482–2491, Osaka, Japan, Dec. 2016. ACL.
25) P. Nakov, Z. Kozareva, S. Rosenthal, V. Stoyanov, A. Ritter, and T. Wilson. SemEval-2013 Task 2:
Sentiment analysis in Twitter. In Second Joint Conference on Lexical and Computational Semantics
(*SEM), Volume 2: Proceedings of the 7th International Workshop on Semantic Evaluation, pages
312–320, Atlanta, Georgia, June 2013. ACL.
26) J. Barnes, R. Klinger, and S. Schulte im Walde. Assessing state-of-the-art sentiment models on state-
of-the-art sentiment datasets. In Proceedings of the 8th Workshop on Computational Approaches to
Subjectivity, Sentiment and Social Media Analysis, pages 2–12, Copenhagen, Denmark, Sept. 2017.
Association for Computational Linguistics.

SENTIMENT ANALYSIS OF BANGLA-ENGLISH CODE-MIXED AND TRANSLITERATED SOCIAL MEDIA COMMENTS USING MACHINE LEARNING

More Related Content

Similar to SENTIMENT ANALYSIS OF BANGLA-ENGLISH CODE-MIXED AND TRANSLITERATED SOCIAL MEDIA COMMENTS USING MACHINE LEARNING (20)

More from indexPub (20)

Recently uploaded (20)

SENTIMENT ANALYSIS OF BANGLA-ENGLISH CODE-MIXED AND TRANSLITERATED SOCIAL MEDIA COMMENTS USING MACHINE LEARNING