Sentiment Analysis in Analysing Monkeypox-Related Tweets Based

on Deep Learning

Yidan Wang

International College, Jinan University, Guangdong, China

Keywords: Sentiment Analysis, Deep Learning, Social Media, CNN-LSTM, Natural Language Processing.

Abstract: As the internet and social media platforms have rapidly expanded, sentiment analysis has emerged as a

significant branch of natural language processing, focused on understanding individuals' emotions and

attitudes toward specific topics. This article provides a comprehensive review of sentiment analysis evolution,

from early dictionary-based methods to modern deep learning techniques. The focus is on a comparative

evaluation of model outcomes from particular research, underscoring the effective performance of a combined

Convolutional Neural Networks- Long Short-Term Memory (CNN-LSTM) deep learning model in analyzing

sentiment within Monkeypox-related tweets. This model harnesses the local feature recognition of CNNs and

the sequential data processing of LSTMs for accurate sentiment detection. Extensive experiments have

demonstrated that this model outperforms standalone CNN-LSTM models in terms of stability and

generalization capabilities. Future research will focus on utilizing more sophisticated sentiment analysis

techniques, such as hierarchical attention networks, and cross-domain models, to enhance precision and

applicability in various practical applications.

1 INTRODUCTION

Sentiment analysis also referred to as opinion mining,

falls under the domain of natural language

processing. It aims to analyze and understand the

feelings, emotions, opinions, and attitudes that people

express regarding a specific topic or subject (Medhat,

2014). It has garnered significant attention in recent

years, particularly in fields such as customer

satisfaction analysis and monitoring patients' mental

health (Wankhade, 2022). This surge in interest can

be attributed to the diverse array of data sources

available, as well as advancements in technologies

such as blockchain, cloud computing, and the Internet

of Things (IoT). The effectiveness of sentiment

analysis is further bolstered by the availability of

sentiment dictionaries and corpora, which offer rich

but varied resources for analysis. Utilizing sentiment

analysis allows companies to discern customer

emotions and opinions shared on social media,

enhancing their understanding of consumer

preferences and behaviors. Consequently, this insight

supports strategic decision-making, influencing

marketing tactics, product improvement, and swift

https://orcid.org/0009-0008-9273-9756

resolution of emerging issues. Ultimately, sentiment

analysis contributes to improving brand

competitiveness and enhancing consumer satisfaction

(Yang, 2020).

Sentiment dictionaries were the initial approach to

sentiment analysis that relied on analyzing the

quantity and polarity of emotional words in a text.

These dictionaries were created either manually or

through machine learning, and analyses that used

rule-based or statistical techniques such as Naïve

Bayes and Support Vector Machines were performed

(Turney, 2002; Pang, 2002). While sentiment

dictionary-based approaches efficiently capture

unstructured features of texts, and are easy to analyze

and understand, these methods need consistent

updating since idioms and internet-specific language

have continuously evolved due to the fast evolution

of the internet and rapid information updates, plus

they face polysemy issues. Traditional machine

learning-based sentiment classification methods

focus on extracting emotional features and selecting

an appropriate combination of classifiers; thus

making a significant impact on the analysis results.

Despite this, these methods frequently overlook the

Wang, Y.

Sentiment Analysis in Analyzing Monkeypox-Related Tweets Based on Deep Learning.

DOI: 10.5220/0012923200004508

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 1st International Conference on Engineering Management, Information Technology and Intelligence (EMITI 2024), pages 215-220

ISBN: 978-989-758-713-9

215

contextual semantics of the text, which has a

detrimental effect on classification accuracy.

Sentiment analysis leveraging deep learning utilizes

methods like Convolutional Neural Networks (CNN),

Recurrent Neural Networks (RNN), and Long Short-

Term Memory (LSTM) networks. The study

categorizes deep learning for sentiment analysis into

four types: analysis using a single neural network,

combined neural network models, analysis

incorporating attention mechanisms, and the

application of pre-trained models (Zhang, 2018;

Yadav, 2020). Furthermore, it suggests that

forthcoming sentiment analysis research should

emphasize the integration of multimodal data,

employing deep learning for cross-domain and real-

time sentiment analysis, and conducting cross-

cultural studies to improve the depth, precision, and

practicality of sentiment interpretation across various

sectors (Tan, 2023; Kaur, 2022; Zadeh, 2017).

The objective of this research is to deliver an

exhaustive examination of sentiment analysis,

spanning from the initial acquisition of text data

through to the various phases of its processing. It will

delve into the mainstream sentiment analysis methods

throughout different historical periods, summarizing

and categorizing these approaches. Additionally, it

explores the various domains and applications of

sentiment analysis, while also addressing the current

research bottlenecks and future directions. The

primary aim of this detailed review is to furnish

insights for experts in the domain and to equip

novices with essential knowledge and guidance.

The layout of the paper is divided into four parts.

The introduction outlines the importance and scope of

emotional analysis, laying the foundation for

subsequent discussions. Next, this paper elaborates

on key terms and concepts related to emotional

analysis. It delves into the various sources of text data

used in sentiment analysis, as well as the

preprocessing steps required for effective analysis.

The third section provides readers with a temporal

perspective on the evolution of this field. This section

explores different fields of practicality in sentiment

analysis, showcasing the practical performance and

significance of its core methods. Discuss the existing

challenges in emotional analysis research and

propose potential avenues for future exploration and

development in this field. Finally, the summary will

highlight the main discoveries and insights from the

analysis, underlining the significance of sentiment

analysis and its anticipated progression in both

academic and practical arenas. Through these

structured sections, readers will gain a comprehensive

understanding of emotional analysis, its methods,

applications, and future prospects.

2 METHODOLOGIES

2.1 Dataset Description and

Preprocessing

Sentiment analysis research relies on key datasets like

the Stanford Sentiment Treebank (SST), the Large

Movie Review Dataset, and the Quora Question Pairs

(QQP) (Zadeh, 2017). These datasets are pivotal for

developing and testing sentiment analysis methods,

offering diverse text samples, and enabling the

exploration of different strategies. Two primary

approaches are dictionary-based and machine

learning-based methods. Dictionary-based techniques

use predefined word lists with sentiment values to

assess text sentiment across various contexts, while

machine learning strategies involve training on these

datasets to achieve deeper, context-sensitive

sentiment understanding. These datasets provide

critical platforms for advancing sentiment analysis

techniques, accommodating various textual analyses,

and facilitating continuous improvement of

methodologies.

2.2 Proposed Approach

This study offers a comprehensive overview of

sentiment analysis, charting its evolutionary path,

examining its methodologies, exploring diverse

applications, and pinpointing future research

directions. To ensure clarity and focus, the study

emphasizes specific research objectives and

methodological frameworks. The author introduces

the fundamental technology underpinning sentiment

analysis, elucidating its core concepts and pivotal

modules. A crucial aspect of this approach involves

visually depicting the sentiment analysis pipeline,

illustrating the sequential steps from data acquisition

and preprocessing to the application of sentiment

analysis techniques and the interpretation of findings.

By delineating this systematic process, the study not

only elucidates the study's aims but also provides a

structured roadmap for conducting sentiment analysis

research, fostering a deeper comprehension of its

applications and potential advancements.

Furthermore, the paper highlights three prominent

methods: CNNs, RNNs, and LSTM networks. Figure

1 illustrates these concepts.

EMITI 2024 - International Conference on Engineering Management, Information Technology and Intelligence

216

Figure 1: The pipeline of the model

(Photo/Picture credit: Original).

2.2.1 Sentiment Lexicon

The analysis method based on sentiment dictionaries

is one approach to sentiment mining and analysis. The

general procedure involves: first, matching the text

with sentiment words; then, aggregating and scoring

these sentiment words; and finally, determining the

text's sentiment orientation. Currently, the most

widely used sentiment dictionaries are mainly of two

types: one is the BosonNLP sentiment dictionary, and

the other is a sentiment dictionary introduced by

CNKI. Sentiment analysis via the BosonNLP

sentiment lexicon follows a clear-cut process.

Initially, the text undergoes segmentation into

sentences and individual words, utilizing tools like

jieba for division. Subsequently, this segmented word

array is cross-referenced with the BosonNLP lexicon

to log the sentiment values of corresponding words.

The summation of these values follows, determining

the overall sentiment: a total score above zero

indicates positivity, whereas a score below zero

suggests negativity. This procedure is depicted in

Figure 2.

Figure 2: The process based on the Sentiment lexicon

(Photo/Picture credit: Original).

Machine learning-based analytical techniques

train algorithms on provided datasets to forecast

results, a practice that has proven highly effective.

For sentiment analysis, these methods harness

extensive labeled or unlabeled text corpora, apply

statistical algorithms for feature extraction, and

perform analysis to generate insights. Sentiment

classification leveraging machine learning is broadly

segmented into supervised, semi-supervised, and

unsupervised techniques. Supervised classification

relies on data samples annotated with emotional

polarity to identify sentiment categories, though it

requires extensive manual labeling and processing.

Typical supervised algorithms include k-nearest

neighbor (KNN), Naive Bayes, and Support vector

machine (SVM). Semi-supervised techniques

augment sentiment classification outcomes by

leveraging features from unlabeled texts and help

mitigate the lack of labeled data. Unsupervised

classification assigns sentiments by analyzing textual

similarities, albeit it's a less frequent approach in

sentiment analysis.

Deep learning analysis methods utilize neural

networks for deeper insight extraction. Prominent

neural network architectures for learning include

CNNs, RNNs, and LSTMs. Deep learning for

sentiment analysis is diverse, encompassing: the

utilization of singular neural network models,

integration of multiple neural networks into hybrid

models, incorporation of attention mechanisms for

enhanced focus within models, and application of

pre-trained models for efficiency gains in analysis.

2.2.2 CNNs

Comprising layers of neurons, CNNs differ from fully

connected networks, where each layer's neurons are

interlinked with adjacent ones. In CNNs,

interconnections occur selectively between layers,

forming a three-dimensional structure from the input

to convolutional and pooling layers. Distinct from

fully connected layers, CNNs include an array of

Sentiment Analysis in Analyzing Monkeypox-Related Tweets Based on Deep Learning

217

layers: an initial input layer, several convolutional

and pooling layers, and a fully connected layer,

followed by a SoftMax layer for output in

classification tasks. Feature extraction is conducted

by the convolutional layer using filters, which move

across the input, identifying patterns and local

connections. To condense the feature map while

preserving essential information, the pooling layer

employs techniques like max or average pooling,

which also diminish computational load and help

avert model overfitting. The fully connected layer

synthesizes these features for output analysis,

whereas the SoftMax layer transforms the fully

connected layer's outputs into a probabilistic

distribution for multi-class classification challenges.

In the field of text sentiment analysis, each

component of CNNs offers distinct advantages. The

convolutional layer can effectively identify local

patterns in text, such as fixed collocations or semantic

features between words, crucial for capturing

emotional expressions. The pooling layer, by

reducing the number of parameters and

computational complexity, not only speeds up the

training process but also helps the model abstract key

information from broader text regions, enhancing the

model's adaptability to different text lengths and

structures. The fully connected layer integrates these

features to form a comprehensive judgment of the

text's emotional tendency. The SoftMax layer then

translates this judgment into specific emotional

category probabilities, facilitating accurate

classification. Although CNNs have demonstrated

strong performance in processing visual data and

have shown unique advantages in text sentiment

analysis, they still have limitations. Additionally, the

performance of CNN models largely depends on the

design of the convolutional kernels and parameter

selection, requiring extensive experimentation and

tuning for optimization. Thus, despite CNNs'

significant potential application in text sentiment

analysis, their limitations and applicable scenarios

must be carefully considered in practical applications.

2.2.3 RNN and LSTM

RNNs, resembling typical neural networks, excel in

processing sequential data by retaining contextual

information across time steps. In translation tasks,

maintaining coherence between successive words is

vital, as translating word by word often leads to

inaccuracies. RNNs address this by allowing hidden

layer neurons to interact, preserving output from prior

steps to influence subsequent word translations,

thereby enhancing accuracy and cohesion. However,

RNNs encounter challenges with lengthy sequences

due to vanishing or exploding gradients, impeding

their ability to learn distant dependencies. To mitigate

this, LSTMs were introduced, featuring specialized

gating mechanisms to manage long-term

dependencies effectively. The forget gate is

responsible for deciding which information from the

past should be thrown away, the input gate refreshes

the internal states with pertinent data, and the output

gate controls the flow of important information to be

outputted. LSTMs outperform standard RNNs in

tasks necessitating long-term context consideration.

Additionally, preprocessing techniques like

Word2vec enhance model performance by providing

richer input, enabling LSTMs to process and translate

text data more accurately.

Word2vec learns semantic information by

embedding words into vectors, grouping similar

meanings in a multidimensional space. It maps words

from their original space to a new one, predicting

surrounding words based on initial word vectors and

adjusting them through backpropagation to match

target words. Similar words sharing contexts

converge in this space. Commonly employed are the

CBOW and Skip-gram architectures; CBOW

forecasts a target word based on context words, while

Skip-gram anticipates context words starting from a

solitary target word. The Gensim package facilitates

training, where the corpus is tokenized, stopwords

removed, and words without vectors eliminated.

LSTM, an enhanced RNN (Wang, 2016), integrates

additional gates to control information flow. Figure 3

depicts the entire process.

Figure 3: The whole process of LSTM (Photo/Picture

credit: Original).

3 RESULTS AND DISCUSSION

This section is dedicated to examining the trends in

loss and accuracy for CNN, LSTM, and their

combined CNN-LSTM frameworks. It seeks to

discern the effectiveness, as well as the potential

advantages and limitations, of each model within the

scope of identifying sentiments in text.

3.1 CNN

In this part, the author assesses the CNN model’s

performance in sentiment analysis using a

Monkeypox tweets dataset. Figure 4 displays the

EMITI 2024 - International Conference on Engineering Management, Information Technology and Intelligence

218

model's (Mohbey, 2023) loss and accuracy curves on

both training and test sets. Initially, the training loss

sharply declines and stabilizes after the first 10

epochs. Conversely, the test loss exhibits

fluctuations, particularly between the 30th and 60th

epochs, suggesting potential issues with

generalization or overfitting. While training accuracy

improves rapidly, it eventually plateaus, whereas test

accuracy, albeit increasing, remains consistently

lower and displays minor fluctuations. This

discrepancy implies potential overfitting of the model

to the training data. Given that CNNs are primarily

designed for image data, their application to textual

data presents limitations, possibly overlooking

essential text features and hindering generalization.

Overall, while the CNN model fits well with the

training data, it struggles to generalize to the test data.

To address this, adjustments such as regularization

techniques to mitigate overfitting and data

enrichment may be necessary. Additionally, tailoring

CNN architectures specifically for text data

processing could enhance performance.

Figure 4: Training and Validation Performance of CNN

Model (Mohbey, 2023).

3.2 LSTM

Figure 5 illustrates the training and test loss and

accuracy of the LSTM model. From a loss

perspective, the training loss initially drops sharply

and then stabilizes, while the test loss follows a

similar downward trend but with more pronounced

fluctuations, particularly in later training stages. This

fluctuation indicates a possible lack of stability in the

model's capacity to generalize from new, unobserved

data. The rapid initial decrease in training loss

indicates efficient early learning, but subsequent

spikes may indicate challenges in model

generalization, possibly stemming from imbalanced

data distribution, inadequate parameter tuning, or

inappropriate learning rate settings leading to

gradient explosion.

Regarding accuracy, the training accuracy

steadily increases after an initial rise in the early

epochs. Similarly, the test accuracy mirrors this

pattern in the first 10 epochs but experiences

significant fluctuations thereafter. This inconsistency

suggests a suboptimal generalization of the model,

possibly influenced by parameter settings and the

dynamic learning environment. Therefore, while

LSTM is theoretically well-suited for textual data,

practical application necessitates meticulous model

design, parameter tuning, and validation tailored to

the specific application scenario and dataset.

Figure 5: Training and Validation Performance of LSTM

Model (Mohbey, 2023).

3.3 CNN-LSTM

The study demonstrates that, when evaluating

Monkeypox-related tweet sentiments, the integrated

CNN-LSTM model outperforms its individual CNN

and LSTM counterparts in both features and

effectiveness. Specifically, the CNN-LSTM model

demonstrated greater stability in loss rates (show in

Figure 6), with smoother curves for both training and

test losses, indicating its effectiveness in avoiding

overfitting and showcasing strong generalization

capabilities when encountering unseen data. In terms

of accuracy, the hybrid model not only achieved

higher accuracy during training and testing but also

maintained a smaller gap between the two, reflecting

its consistent predictive power across different types

of data. This enhancement likely stems from the

effective combination of CNN local feature

extraction and LSTM sequential dependency

processing, with each complementing the other to

ensure accuracy and robustness in text sentiment

analysis.

Sentiment Analysis in Analyzing Monkeypox-Related Tweets Based on Deep Learning

219

Figure 6: Training and Validation Performance of CNN-

LSTM Model (Mohbey, 2023).

3.4 Future Extends

Navigating the intricacies of sentiment analysis

requires overcoming several obstacles, including the

deciphering of unstructured or ironic expressions, the

need for more nuanced sentiment categorization,

reliance on data with annotations, limitations inherent

to present word embedding techniques, and biases

embedded in the datasets used for training. It is

imperative that future studies focus on elevating

precision and broadening the scope of application.

This involves the creation of models that adeptly

discern subtle emotional nuances in text, assess the

sentiment variance pertaining to different topics,

adeptly handle texts with ambiguity and sarcasm,

broaden sentiment analysis to encompass a wider

range of languages, and enhance the efficacy of

sentiment analysis on various social media platforms

(Tan, 2023).

Potential approaches may involve employing

hierarchical attention networks for nuanced sentiment

analysis, integrating topic modeling with sentiment

analysis to analyze sentiment distribution, leveraging

reinforcement learning techniques to address sarcasm

and ambiguity, creating cross-lingual models using

transfer learning methods, and generating domain-

specific embeddings tailored for social media texts.

By surmounting these challenges and advancing

model capabilities, sentiment analysis can broaden its

practical applications and provide deeper insights into

textual sentiments.

4 CONCLUSIONS

This paper provides a sample but comprehensive

review of sentiment analysis, covering the entire

spectrum from initial data sourcing to subsequent

processing phases. Through an in-depth examination

of a hybrid CNN-LSTM model, it has been

established that such an approach enhances both

accuracy in sentiment detection and robustness in

sentiment expression. The experimental evaluation

revealed that the CNN-LSTM hybrid model exhibits

superior stability and generalization compared to

standalone CNN or LSTM models. Moving forward,

future research endeavors will focus on refining

sentiment analysis techniques, integrating cross-

linguistic models, and enhancing sentiment analysis

effectiveness on social media platforms.

REFERENCES

Kaur, R., & Kautish, S. 2022. Multimodal sentiment

analysis: A survey and comparison. Research anthology

on implementing sentiment analysis across multiple

disciplines, pp: 1846-1870.

Medhat, W., Hassan, A., & Korashy, H. 2014. Sentiment

analysis algorithms and applications: A survey. Ain

Shams engineering journal, vol. 5(4), pp: 1093-1113.

Mohbey, K. K., Meena, G., Kumar, S., & Lokesh, K. 2023.

A CNN-LSTM-based hybrid deep learning approach

for sentiment analysis on Monkeypox tweets. New

Generation Computing, pp: 1-19.

Pang, B., Lee, L., & Vaithyanathan, S. 2002. Thumbs up?

Sentiment classification using machine learning

techniques. arXiv:0205070.

Tan, K. L., Lee, C. P., & Lim, K. M. 2023. A survey of

sentiment analysis: Approaches, datasets, and future

research. Applied Sciences, vol. 13(7), p: 4550.

Turney, P. D. 2002. Thumbs up or thumbs down? Semantic

orientation applied to unsupervised classification of

reviews. arXiv: 0212032.

Wang, J., Yu, L. C., Lai, K. R., & Zhang, X. 2016.

Dimensional sentiment analysis using a regional CNN-

LSTM model. In Proceedings of the 54th annual

meeting of the association for computational linguistics,

vol. 2, pp: 225-230.

Wankhade, M., Rao, A. C. S., & Kulkarni, C. 2022. A

survey on sentiment analysis methods, applications,

and challenges. Artificial Intelligence Review, vol.

55(7), pp: 5731-5780.

Yadav, A., & Vishwakarma, D. K. 2020. Sentiment

analysis using deep learning architectures: a review.

Artificial

Yang, L., Li, Y., Wang, J., & Sherratt, R. S. 2020.

Sentiment analysis for E-commerce product reviews in

Chinese based on sentiment lexicon and deep learning.

vol. 8, pp: 23522-23530.

Zadeh, A., Chen, M., Poria, S., Cambria, E., & Morency, L.

P. 2017. Tensor fusion network for multimodal

sentiment analysis. arXiv:1707.07250.

Zhang, L., Wang, S., & Liu, B. 2018. Deep learning for

sentiment analysis: A survey. Wiley Interdisciplinary

Reviews: Data Mining and Knowledge Discovery, vol.

8(4), p: e1253.

EMITI 2024 - International Conference on Engineering Management, Information Technology and Intelligence

220