An Overview of Sentiment Analysis: Levels, Approaches and

Challenges

Loukmane Maada

, Khalid Al Fararni

, Badreddine Aghoutane

,Yousef Farhaoui

Mohammed Fattah

IA Laboratory, Science Faculty, Moulay Ismail University, Meknes, Morocco

LISAC Laboratory, Faculty of Sciences Dhar El Mahraz, Sidi Mohamed Ben Abdellah University, Fez, Morocco

L-STI, T-IDMS, University of Moulay Ismail, Faculty of Science and Technics, Errachidia, Morocco

Image laboratory, Moulay Ismail University Meknes, Morocco.

Keywords: Sentiment analysis, machine learning, deep learning, lexicon-based

Abstract: Text is a huge data source; it contains opinions, facts, feelings ... basically embedded knowledge. Sentiment

Analysis (S.A) primary goal is to analyse the text and determine its polarity (positive, negative, neutral). This

field of research has been on the rise since the beginning of this century. A lot of approaches, from word

preprocessing and embedding techniques to complex big data architecture., have been modelled, tested and

proposed. This paper provides an overview of the different sentiment analysis approaches, namely the

traditional machine learning approach, the deep learning approach, the lexicon-based approach, and the hybrid

approach. In addition, a brief insight into the challenges S.A faces and some proposed solutions are displayed.

1 INTRODUCTION

The emergence of Web 2.0 has changed how people

use the internet, from a read-only paradigm to an

interactive paradigm. This digital revolution allowed

the users to express their opinions and feelings about

different topics: politics, finance, health and more;

Especially with the emergence of social media,

micro-blogging websites with millions of users

generate an enormous amount of data per day. For

instance, about 500 million tweets are sent out per

day. This embedded knowledge within social media

drew the attention of researchers; however, the

classical NLP methods were suboptimal, leading to

the development of an NLP sub-field known as

sentiment analysis.

Sentiment analysis, also known as opinion mining

or emotion A.I, is a set of analytic methods that

extract and identify information (sentiment, opinion,

and attitude) from text using natural language

https://orcid.org/0000-0003-4165-1486

https://orcid.org/0000-0001-5907-6948

https://orcid.org/0000-0002-9555-6786

https://orcid.org/0000-0001-6128-9715

https://orcid.org/0000-0003-0870-6262

processing, text analysis, and computational

linguistics. This information is used to analyse

customers ' satisfaction with a product. Or to predict

particular behaviours, e.g. forecasting the election

winner.

Opinion mining has been used in different

domains and multiple real-world applications. From

predicting the stock market (Chiong Raymond et

al.,2018) to detect event popularity (Mariana Daniel

et al.,2016) and predicting the election (F. Nausheen

et al.,2018); to fighting Covid-19 (A.H. Alamoodi et

al., 2020), increasing the benefits of tourism‘s

company (Gianpierre Zapata et al., 2017) and to

improve tourists' experience (L.Maada et al., 2021).

Our aim is to build a big data solution based on

hybrid recommendation and sentiment analysis using

machine and deep learning techniques to recommend

the most suitable tourist offer, improve the customer

experience and forecast the tourist demand in

Morocco (

K. AL Fararni et al., 2021)

Maada, L., Al Fararni, K., Aghoutane, B., Farhaoui, Y. and Fattah, M.

An Overview of Sentiment Analysis: Levels, Approaches and Challenges.

DOI: 10.5220/0010735700003101

In Proceedings of the 2nd International Conference on Big Data, Modelling and Machine Learning (BML 2021), pages 413-418

ISBN: 978-989-758-559-3

413

This paper gives an overview of sentiment

analysis, which is beneficial for newcomer

researchers. It provides a comprehensive summary of

the different sentiment analysis levels(sentence level,

document level, and aspect level). Followed by the

basic approaches of sentiment analysis(machine

learning approach, lexicon-based approach, hybrid

approach, and deep learning approach), then the paper

will expose the challenges sentiment analysis faces

and some proposed solutions.

2 LEVELS OF SENTIMENT

ANALYSIS

Sentiment analysis has been investigated on three

different granularity levels: sentence level, document

level, and aspect level.

2.1 Sentence-level

This level focuses on the sentences in a text and tries

to extract opinions from them. It first classifies the

sentence as either objective or subjective. Then it

extracts the polarity (positive or negative) of each

sentence, assuming that objective sentences are

always neutral. However, objective sentences are not

always neutral and may include some opinions. For

example, "My phone broke yesterday," which

contains an implicit negative sentiment (Singh Nikhil

Kumar et al., 2020); we refer to this sort of sentiment

as a mild sentiment (Zimba David et al., 2018).

2.2 Document-level

At this level, we suppose that the whole document

contains one opinion; in other words, it assumes that

each document has a single entity opinion. Hence,

Sentiment analysis at this level does not apply to

documents that contain multiple entities (Zimba

David et al., 2018).

2.3 Aspect-level

This level is considered to be the hardest out of the

three levels. It assumes that every opinion is

dependent on an aspect and has a target which without

it is meaningless (Zimba David et al., 2018).

Therefore the aspect level will aim to extract the

entities and their respective aspects. To illustrate the

sentence "The PS5 design is great, but it's costly",

evaluate two parts: the great design and the high

price.

3 SENTIMENT ANALYSIS

APPROACHES

In this section, we will display the different

approaches to sentiment analysis. This section will be

divided into four subsections: the lexicon-based

approach, the machine learning approach, the hybrid

approach, and the deep learning approach. Each of

those subsections provides an overview of different

techniques that those approaches use.

3.1 Lexicon-based Approach

The lexicon-based approach is based on using a

corpus or dictionary to determine the polarity. This

approach requires a high-quality sentiment lexicon to

yield good results. There are mainly two types of

lexicons: The general-purpose lexicon which, as its

name suggests, is used for general-purpose

classifiers, and the domain-specific lexicon used for

the domain-related classifiers. In this section, we will

display the lexicon-based approaches and some of the

most used worldwide lexicons.

3.1.1 Manual Approach

This approach utilises an existing dictionary or corpus,

then associates the sentiment strength to each of the

sentiments. For example, if we consider a 3-class

sentiment strength {-1,0,1} the words "good" and

"bad" will be associated respectively with the

sentiment strength 1 and -1. The lexicon's advantages

are that it works well and is quite accurate. In contrast,

this method has some drawbacks. It is very time-

consuming, cannot match the language's evolution,

and is not qualified for specific domains. Generally

speaking, rather than its direct use. The manually

developed lexicon is used to blend with other

methods.

3.1.2 Dictionary Approach

The dictionary approach was first presented by (Hu

Minqing et al., 2004). It consists of building a small

dictionary with a few sentiment words whose polarity

has been manually set. The number of words is

increased iteratively by gathering antonyms and

synonyms from well-known lexicons. e.g. WordNet

or its enhanced version SentiWordNet.

3.1.3 Corpus Approach

The corpus approach, like the dictionary approach,

starts with a small set of words;

BML 2021 - INTERNATIONAL CONFERENCE ON BIG DATA, MODELLING AND MACHINE LEARNING (BML’21)

414

then increases among the large corpus according to

some specific rules or formulas. The major difference

between the two methods is that the corpus approach

can find opinion words within a particular context.

There are two types of corpus-based techniques:

statistics-based (PMI (Kenneth Ward Church et al.,

1989), LSA (Thomas K Laudauer et al., 1998)...) and

semantic-based approaches when it comes to

generating techniques.

3.1.4 Lexicon Examples

WorldNet: is a cognitive linguistics-based English

dictionary suggested by (Miller et al.,1990). It

associates words with semantic relations such as

synonyms, hyponyms, and meronyms. The synonyms

are organised into synsets, each with a brief definition

and usage examples.

SentiWorldNet: is a lexical resource for opinion

mining based on WordNet. It classifies the word into

positive, negative, and neutral. SentiWordNet takes

into consideration that a world may have different

polarities. For instance, ”cool” may mean having a

low temperature as in a cool afternoon or calmness as

in he recovered his cool. SentiWord uses glosses for

each word entry to distinguish one from another

(Esuli Andrea et al., 2006).

WordNet-Affect: is a hand-curated collection of

emotion-related words (nouns, verbs, adjectives, and

adverbs), divided into 28 subcategories (“Joy”,

“Love”, “Fear”, etc.) and classified as “Positive”,

“Negative”, “Neutral”, or “Ambiguous”.

3.2 Machine Learning Approach

The machine learning approaches consist of two

categories that are supervised learning and

unsupervised learning. In this section, we will be

exposing the widely used algorithms in both

categories.

3.2.1 Supervised Learning

Supervised learning is a set of learning algorithms

trained based on labelled data. In the following, we

will briefly expound on the widely used techniques in

sentiment analysis.

Naive Bayes is a probabilistic model based on the

Bayes formula Eqs.1 with an independence

assumption between the entities(features).

𝑃𝐴/𝐵 

∗/



(1)

The Naive Bayes (N.B) or simple Bayes is a fast,

robust, and simple method that is patriotically

successful in NLP.( Fitri et al., 2018) studied the

customer's satisfaction with cellular data service

using N.B algorithm on a Tweeter-based data set.

They reached on average 94,5% precision, 93,31%

recall, F1-score of 93,15% and accuracy of 99,09%.

(Risky Novendri et al., 2020) worked on analyzing

YouTube comments using the N.B method with

which they achieved 81%, 74.83%, and 75.22% for

accuracy, precision, and recall, respectively.

Support Vector Machine or SVM is a binary

classification method that determines the optimal

hyperplane separating the data to positive and

negative in S.A. The SVM proceeds by maximising

the margin between the training data. If the training

set is not linearly separable, it is mapped to a higher

dimension using the kernel function. (Lavanya and

Deisy, 2017) present a multi-class SVM model to

analyse customer's satisfaction with the

company's(Google, Apple, Microsoft...) products

using Tweets. The model reached an accuracy of

96%, a recall of 83%, and an F1-score of 88%. In the

tourism field, ( Laoh et al., 2019) used SVM with n-

gram to examine the Bali hotel review, they achieved

the best result using the 2-gram method combined

with SVM, 94%, 75%, 72%, 74% for accuracy,

precision, recall, and F1-measure respectively.

Maximum entropy is a probability distribution

technique widely used in NLP. It consists of finding

the distribution that satisfies the constraints and

should have the maximum entropy. Consider the

following data set T = {(x

, y

), (x

, y

}),..., (x

, y

)}

and feature function f

(x, y), i = 1, 2,..., n. The

equations that make up the model are as follows:

𝑝

/





∑









,









(2)

where 𝜆



is the weight parameter of the feature

function and

𝑍





∑

𝑒𝑥𝑝

∑

𝜆



𝑓



𝑥, 𝑦







(3)

(Dewanti Putri et al., 2019) led a study on the Grab

user reviews in which they utilised the SVM

algorithm and the MaxE algorithm. The results were

close, with the SVM reaching a maximum accuracy

of 89%, whereas MaxE reached 90.46%.

3.2.2 Unsupervised Learning

Unlike supervised learning that heavily relies on

labelled data, unsupervised learning methods detect

patterns in the data set without pre-associated labels.

An Overview of Sentiment Analysis: Levels, Approaches and Challenges

415

K-means clustering is a method that aims to

partition the data set into a k cluster in which each

data point belongs to the cluster with the nearest

mean. (Riaz et al., 2019) used the K-means clustering

to partition the words sentiment strength into three

clusters, positive, negative, and neutral. The review

analysis technique they proposed reached an accuracy

of 95%.

Topic modelling is an unsupervised machine

learning approach that scans the text, detects the

words and sentence patterns, clusters word groups

and similar expressions that best describes a set of

texts.

(K. Lavanya et al., 2018) suggested a topic

modelling-based framework to do multi-aspect S.A

on Chinese Online social reviews. They used the

Latent Dirichlet Allocation (LDA) to extract the

global topic, the local topic (the aspect) and their

associated sentiment. The empirical results revealed

that the proposed approach enhanced the performance

of the model.

3.3 Deep Learning Approach

Deep learning approaches are well known for trying

to mimic the human's brain work. The deep learning

algorithms were mostly used in image recognition,

but with the embedding tools (word2vec and GloVe)

emerging, deep learning methods become the

research focus in the sentiment field. In this section,

we will display a few deep learning algorithms.

3.3.1 Recurrent Neural Network

A recurrent neural network (R.N.N) is a neural

network algorithm derived from feedforward neural

networks. R.N.N passes information from its

previous time step to the current time step, making it

capable of processing sequential data like text.

3.3.2 Long Short-Term Memory Model

L.S.T.M model is an extension of the R.N.N model

that utilises gates to ease the issues of blow up and

vanishing gradients in the R.N.N model. The

L.S.T.M, unlike the R.N.N model, discards the

useless information (forget gate) and passes the useful

one to the next unit.

(Fu et al., 2018) proposed a lexicon-enhanced

L.S.T.M model. They trained a word sentiment

classifier with a sentiment lexicon to obtain the

embedded sentiment of each word, the word

embedding, and its relative sentiment is passed as an

input to L.S.T.M. The results show that this method

enhances the performance. For instance, it improved

the accuracy by about 4% on the IMDB dataset.

3.3.3 Convolutional Neural Network

A deep convolutional neural network (D.C.N.N) is

mostly applied in the field of visual imagery analysis.

However, it has high efficiency when combined with

good word representation models.

(Minaee et al., 2019) proposed a framework that

combines L.S.T.M and C.N.N models using GloVe

embedding representation. They average the

prediction from L.S.T.M and C.N.N to get the final

prediction. The experience showed a slight gain in

accuracy compared to the use of CNN and L.S.T.M

independently. For instance, the proposed model

achieved an accuracy of 90% on the IMDB data set,

whereas the CNN and L.S.T.M alone reached 89.3%

and 89%, respectively.

Both CNN and L.S.T.M are one of the most used

neural network approaches in the field of sentiment

analysis. However, each one of them does a better job

at a specific task. C.N.N excels in features extraction,

whereas L.S.T.M achieves superior sentiment

understanding, hence the combination of both in

recent research.

3.4 Hybrid Approach

The hybrid approach is a relatively new approach to

sentiment analysis that combines different

approaches. It uses each technique's strengths. This

method is typically used for complex sentiment

analysis tasks.

(Asghar et al., 2018) developed a hybrid

technique that incorporates four classifiers: a slang

classifier, an emoticon classifier, the SentiWordNet

classifier, and an enhanced domain-specific

classifier; to improve the performance of Twitter-

based sentiment analysis systems. The sentiment

classification was performed at the sentence level; the

multistage hybrid framework overcame the

limitations of previously proposed models.

(Yadav et al., 2019) compared the performance of

a hybrid based approach that uses SVM, N.B, and

Genetic algorithm (G.A) and each of those algorithms

performance on a hotel reviews data set. The hybrid

approach achieved a 93% accuracy, whereas SVM

alone achieved 85.2%, N.B 85%, and G.A 85.3%.

(Sosa, 2017) combined two neural network

models namely, C.N.N and L.S.T.M, in a hybrid

model. Experiment results showed that the L.S.T.M-

C.N.N model performed better than the average

model.

BML 2021 - INTERNATIONAL CONFERENCE ON BIG DATA, MODELLING AND MACHINE LEARNING (BML’21)

416

The hybrid approach has shown to be the most

effective approach among all existing ones since it

combines different approaches to overcome each

one's weaknesses and makes the most of their

strengths.

4 CHALLENGES

Text is a complicated sequential structure that

includes information, making sentiment analysis

difficult. In this section, we will explore some of

those challenges and some proposed solutions.

4.1 Sarcasm

Sarcasm can be defined as the use of irony to mock or

annoy for a humorous purpose. For instance, "The

movie plot was so good that I knew who the killer was

in the first 5 min :)" this sentence contains first the

word "good" and second the happy face emoji; this

review has a high probability of being classified as

positive whereas it is negative. Much research has

been done to detect sarcastic reviews. The state of art

performance is about 80% accuracy (N.Majumder,

2019).

4.2 Suggestions, Questions and Advice

Suggestions, questions, and advice are mistaken for

either positive or negative sentiment, whereas they

should be classified as neutral. For instance, "It would

be great if the store installs a better security system",

this review is an objective suggestion that is mistaken

for a positive sentiment since it contains "better" and

"great", which are strong positive words.

4.3 Negation

Negation is a common language technique for

expressing displeasure. This technique makes

deducing the meaning of narrative content harder. For

instance, "I'm not happy with the quality of the

product" this review is negative; however, it has a

high probability of getting mistaken for positive.

Some research has been done in this field; (Pröllochs

et al., 2020) proposed a framework based on a

reinforcement learning model that detects the parts of

the sentence that use negation, this method yields

some promising results.

5 CONCLUSION

In this paper, we investigated the levels of sentiment

analysis (aspect-level, document-level, sentence-

level). We presented the different approaches used in

opinion mining: the machine learning approach, the

lexicon-based approach, the deep learning approach

and finally, the hybrid approach. We listed some of

the challenges sentiment analysis faces. We believe

that sentiment analysis techniques are on a constant

rise; and will offer a better and deeper understanding

of complex sentiments in the incoming years.

REFERENCES

Chiong, Raymond and Fan, Zongwen and Hu, Zhongyi and

Adam, Marc T. P. and Lutz, Bernhard and Neumann,

Dirk, 2018 A Sentiment Analysis-Based Machine

Learning Approach for Financial Market Prediction via

News Disclosures. In Proceedings of the Genetic and

Evolutionary Computation Conference Companion,

page 278–279.

Daniel, Mariana and Neves, Rui Ferreira and Horta, Nuno,

2017, Company event popularity for financial markets

using Twitter and sentiment analysis. In Expert Systems

with Applications, volume 71, page 111–124.

F. Nausheen and S. H. Begum, 2018, Sentiment analysis to

predict election results using Python. In 2018 2nd

International Conference on Inventive Systems and

Control (ICISC), page 1259–1262.

Alamoodi, A.H.and Zaidan, B.B. and Zaidan, A.A. and

Albahri, O.S. and Mohammed, K.I. and Malik, R.Q.

and Almahdi, E.M. and Chyad, M.A. and Tareq, Z. and

Albahri, A.S. and Hameed, Hamsa and Alaa, Musaab,

2021, Sentiment analysis and its applications in fighting

COVID-19 and infectious diseases: A systematic

review. In Expert Systems with Applications, volume

167, page 114155.

Gianpierre Zapata. and Javier Murga. and Carlos

Raymundo. and Jose Alvarez. and Francisco

Dominguez, 2017, Predictive Model based on

Sentiment Analysis for Peruvian SMEs in the

Sustainable Tourist Sector. In Proceedings of the 9th

International Joint Conference on Knowledge

Discovery, Knowledge Engineering and Knowledge

Management - KMIS, page 232-240.

David Zimbra and Ahmed Abbasi and Daniel Zeng and

Hsinchun Chen, 2018 The State-of-the-Art in Twitter

Sentiment Analysis: A Review and Benchmark

Evaluation. In ACM Trans. Manage. Inf. Syst, Article 5.

Singh, Nikhil Kumar and Tomar, Deepak Singh and

Sangaiah, Arun Kumar, 2020 Sentiment analysis: a

review and comparative analysis over social media. In

Journal of Ambient Intelligence and Humanized

Computing, volume 11, page 97-117.

An Overview of Sentiment Analysis: Levels, Approaches and Challenges

417

Minqing Hu and Bing Liu, 2004 Mining and summarizing

customer reviews. In Proceedings of the tenth ACM

SIGKDD international conference, page 168-177.

Kenneth Ward Church and Patrick Hanks, 1989 Word

association norms, mutual information, and

lexicography. In Proceedings of the 27th annual

meeting on Association for Computational Linguistics,

page 76–83.

Thomas K Landauer and Peter W. Foltz and Darrell Laham,

1998 An introduction to latent semantic analysis. In

Discourse Processes, volume 25, page 259–284.

Miller, George A. and Beckwith, Richard and Fellbaum,

Christiane and Gross, Derek and Miller, Katherine J.,

1990 Introduction to WordNet: An On-line Lexical

Database*. In International Journal of Lexicography,

volume 3, page 235–244.

Esuli, Andrea and Sebastiani, Fabrizio, 2006

SENTIWORDNET : A Publicly Available Lexical

Resource for Opinion Mining.Proceedings of the Fifth

International Conference on Language Resources and

Evaluation.

F. S. Fitri and M. N. S. Si and C. Setianingsih, 2018

Sentiment Analysis on the Level of Customer

Satisfaction to Data Cellular Services Using the Naive

Bayes Classifier Algorithm. In 2018 IEEE

International Conference on Internet of Things and

Intelligence System (IOTAIS), page 201–206.

Novendri Risky and Callista, Annisa Syafarani and

Pratama, Danny Naufal and Puspita, Chika Enggar,

2020 Sentiment Analysis of YouTube Movie Trailer

Comments Using Naïve Bayes. In Bulletin of

Computer Science and Electrical Engineering, volume

1, page 26–32.

K. Lavanya and C. Deisy, 2017 Twitter sentiment analysis

using multi-class SVM. In International Conference on

Intelligent Computing and Control (I2C2), page 1–6.

E. Laoh and I. Surjandari and N. I. Prabaningtyas, 2019

Enhancing Hospitality Sentiment Reviews Analysis

Performance using SVM N-Grams Method. In 2019

16th International Conference on Service Systems and

Service Management (ICSSSM), page 1–5.

K. B. A. Dewanti Putri and A. Uswatun Khasanah and A.

Azzam, 2019 Sentiment Analysis on Grab User

Reviews Using Support Vector Machine and Maximum

Entropy Methods. In 2019 International Conference on

Information and Communications Technology

(ICOIACT), page 468-473.

Riaz, Sumbal and Fatima, Mehvish and Kamran, M. and

Nisar, M. Wasif, 2019 Opinion mining on large scale

data using sentiment analysis and k-means clustering.

In Cluster Computing, volume 22, page 7149-7164.

K. Lavanya and C. Deisy, 2012 Multi-aspect sentiment

analysis for Chinese online social reviews based on

topic modeling and HowNet lexicon. In Knowledge-

Based Systems, volume 37, page 186-195.

X. Fu and J. Yang and J. Li and M. Fang and H. Wang Bo

Pang and Lillian Lee, 2018 Lexicon-Enhanced LSTM

With Attention for General Sentiment Analysis. In

IEEE Access, volume 6, page 71884-71891.

Shervin Minaee and Elham Azimi and AmirAli

Abdolrashidi, 2019 Deep-Sentiment: Sentiment

Analysis Using Ensemble of CNN and Bi-LSTM

Models.

Asghar, Muhammad Zubair and Kundi, Fazal Masud and

Ahmad, Shakeel and Khan, Aurangzeb and Khan,

Furqan, 2018 MT-SAF: Twitter sentiment analysis

framework using a hybrid classification scheme. In MT-

SAF: Twitter sentiment analysis framework using a

hybrid classification scheme, volume 35, page e12233.

N. Yadav and R. Kumar and B. Gour and A. U. Khan, 2019

Extraction-Based Text Summarization and Sentiment

Analysis of Online Reviews Using Hybrid

Classification Method. In 2019 Sixteenth International

Conference on Wireless and Optical Communication

Networks (WOCN), page 1-6.

P. M. Sosa, 2017 MT-SAF: Twitter sentiment analysis

using combined LSTM-CNN models.

N. Majumder and S. Poria and H. Peng and N. Chhaya and

E. Cambria and A. Gelbukh, 2019 Sentiment and

Sarcasm Classification With Multitask Learning. In

IEEE Intelligent Systems, volume 34, page 38-43.

L. MAADA, K. AL FARARNI, B. AGHOUTANE, M.

FATTAH, Y. FARHAOUI. A proof of concept web

application for sentiment analysis in tourism in the

region of Draa-Tafilalet, presented at the 2nd Edition of

International Conference on Big Data, Modelling and

Machine learning (BML’21), Kenitra, Morocco, July.

15 – 16, 2021.

K. AL Fararni, F. Nafis, B. Aghoutane, A. Yahyaouy, J.

Riffi, A. Sabri. Hybrid Recommender System for

Tourism Based on Big Data and AI: A Conceptual

Framework. Big Data Mining and Analytics 2021, 4(1):

47-55

Nicolas Pröllochs and Stefan Feuerriegel and Bernhard

Lutz and Dirk Neumann, 2020 MT-SAF:Negation

scope detection for sentiment analysis: A reinforcement

learning framework for replicating human

interpretations. In Information Sciences, volume 536,

page 205-221.

BML 2021 - INTERNATIONAL CONFERENCE ON BIG DATA, MODELLING AND MACHINE LEARNING (BML’21)

418