Analysing the Sentiments in Online Reviews with Special Focus on

Automobile Market

Ayman Yafoz

, Farial Syed

Malek Mouhoub

and Lisa Fan

Department of Information Systems, King Abdulaziz University, Abdullah Sulayman Street, Jeddah, Saudi Arabia

Department of Computer Science, University of Regina, 3737 Wascana Parkway, Regina, Canada

Keywords: Canadian Automobiles Market, Sentiment Analysis, Machine Learning, Deep Learning.

Abstract: Analysing the sentiments in online reviews assists in understanding customers’ satisfaction with a provided

service or product, which gives the industry an opportunity to enhance the quality of their commodity, increase

sales volume, develop marketing strategies, improve response to customers, promote customer satisfaction,

and enhance the industry image. However, the studies focusing on applying machine learning algorithms and

word embedding models, as well as deep learning techniques to classify the sentiments in reviews extracted

from automobile forums, are arguably limited, and to fill this gap, this research addressed this area. Moreover,

the research concentrated on categorizing positive, negative, and mixed sentiment categories in online forum

reviews. The procedures for gathering and preparing the dataset are illustrated in this research. To perform

the classification task, a set of models which include supervised machine learning, deep learning, and BERT

word embedding is adopted in this research. The results show that the combination of the BERT word

embedding model with the LSTM model produced the highest F1 score. Finally, the paper lays out

recommendations to enhance the proposed system in future studies.

1 INTRODUCTION

Sentiment analysis has recently gained an increasing

amount of attention from researchers due to its

importance as well as the growth of social media.

However, the current contributions addressing

sentiment analysis on online reviews about

automobiles are arguably not adequate (Wijnhoven et

al., 2017). Furthermore, many phrases that appear in

automobile reviews are solely related to the

automobile industry and are not frequently used in

other reviews. For example, the phrase “It has more

ground clearance than majority of CUVs.” conveys a

positive sentiment, while the phrase “Massive

Hyundai Engine Recall” conveys a negative

sentiment. These factors motivated us to conduct this

work and also to limit the scope of this research to

online automobile reviews.

Moreover, this research is targeting the analysis of

the sentiments into positive, negative, or mixed

categories to provide a fine-grained classification of

sentiments beyond the classical coarse-grained

classification that is limited to only negative and

https://orcid.org/0000-0001-7381-1064

positive sentiments. The data in this research was

gathered from a Canadian online forum specializing

in automobile topics called Autos.ca (Autos.ca). This

work addresses the lack of a customized sentiment

analyser working on text exclusively about Canadian

automobiles with reviews written mostly in Canadian

English. The dataset has been uploaded on GitHub

(named English Automobile Dataset) and can be

freely accessed by future researchers, who only

intend to use it for academic purposes, through the

link in (English Automobiles Dataset).

The code was written in Python due to its large

community, ease of use, libraries (for instance,

Pandas, NLTK, and Sklearn), and the language-

adequate handling of sentiment analysis tasks. On the

other hand, MySQL Server was utilized to save the

dataset as it is compatible with both the Python

environment and Windows operating system.

Moreover, an NVIDIA Tesla P100 GPU with Google

Cloud was utilized to train the deep learning and word

embedding models. The Keras neural networks API

and TensorFlow platform were utilized to provide the

deep learning libraries.

Yafoz, A., Syed, F., Mouhoub, M. and Fan, L.

Analysing the Sentiments in Online Reviews with Special Focus on Automobile Market.

DOI: 10.5220/0010812100003116

In Proceedings of the 14th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2022) - Volume 3, pages 261-267

ISBN: 978-989-758-547-0; ISSN: 2184-433X

261

The remaining sections of this paper describe the

different steps of our proposed methodology and are

divided as follows. Section 2 provides a literature

review discussing several sentiment analysis

contributions. Section 3 explains in detail the data

assembly and annotation phases. Section 4

highlights the phases of the pre-processing, while

Section 5 shows the phases of the feature selection.

Section 6 illustrates the phases of splitting and

balancing the dataset. Section 7 lists and analyses

the outcomes of the machine learning models.

Section 8 overviews the BERT word embedding

model as well as the deep learning models, and

reports on the related experimental results. Section

9 focuses on the data visualization technique.

Finally, Section 10 provides a list of concluding

remarks and future enhancements to improve the

quality of this research work.

2 LITERATURE REVIEW

(Yafoz et al., 2021) analysed the sentiments in

datasets containing Arabic online reviews about

Arabic real estate and automobiles. They employed

the BERT word embedding model with a set of deep

learning algorithms: BiLSTM (Bidirectional Long

Short-Term Memory), LSTM (Long Short-Term

Memory), GRU (Gated Recurrent Unit), CNN

(Convolutional Neural Networks), and CNN-GRU.

The automobile dataset had almost 6,585 opinions,

while the real estate dataset contained almost 6,434

opinions. The records in both datasets were split into

three sentiment types (negative, positive, and

mixed). For the dataset concerning automobiles,

the highest F1 score was 98.71% using the BERT

model with the LSTM. On the other hand,

the highest obtained F1 score for the real estate

dataset was 98.67% using the BERT model with the

CNN.

(Malik et al., 2018) performed a sentiment

analysis on a dataset that had 2,000 reviews divided

evenly between positive and negative reviews.

These reviews were written in Roman Urdu

about automobiles. For the classification process,

eight classifiers were utilized: bagging,

Multinomial Naïve Bayes, AdaBoost, Random

Forest, SVM, Deep Neural Network, Decision Tree,

and K Nearest Neighbor. The highest accuracy

result achieved by the Multinomial Naïve Bayes

classifier was 89.75%.

(Alsawalqah et al., 2015) classified the

sentiments in automobile tweets concerning three

automobile manufacturers (BMW, Audi, and

Mercedes). The dataset contained 3000 tweets

divided equally among the three automobile

manufacturers (1000 tweets for each automobile

manufacturer). The researchers used the Naïve

Bayes algorithm to classify the sentiments. The

results reflected that Audi had the lowest negative

polarity (only 16%) and the highest positive polarity

(around 83%) compared with Mercedes and BMW.

This reflected that the reviewers were more satisfied

with Audi than with Mercedes and BMW.

Finally, (Yafoz et al., 2020) classified three

sentiment categories (mixed, positive and negative)

in Arabic online reviews concerning automobiles

and real estate. The dataset of real estate included

6,434 opinions, while the automobile dataset

included 6,585 opinions. The researchers applied

twenty-two machine learning, four word embedding

algorithms (Fasttext, Glove, CBOW, and Skip-

gram), and four deep learning models (LSTM, GRU,

CNN, and BiLSTM) to classify the reviews. For the

Arabic automobile dataset, the highest F1 score was

84.90% by the CBOW with the GRU. On the other

hand, for the dataset concerning real estate, the

highest F1 score was 71.33% with the combination

of the Skip-gram with the GRU and CNN.

3 DATA ASSEMBLY AND

ANNOTATION

The purpose of assembling and labelling the records

of the dataset is to create a dataset that is suitable for

supervised classification.

3.1 Data Assembly

The Octoparse web crawler (Octoparse Web

Scraper) was used to extract and organize the

data due to its simplicity, quickness, efficiency,

and ability to extract and organize the data. The

dataset source is Autos.ca, which is a Canadian

automobile online forum. The dataset size is of 4014

unique records where all duplicated records were

removed using the option of remove duplicates

offered by Microsoft Excel. The dataset was divided

into 2078 positive, 1467 negative, and 469 mixed-

opinion records, and it focused on almost 56

domains related to automobiles. Table 1 shows an

example of analysing the sentiments in reviews from

the dataset.

ICAART 2022 - 14th International Conference on Agents and Artiﬁcial Intelligence

262

Table 1: An Example of a Sentiment Analysis Performed

on the Dataset.

Review Sentiment

I do not like the Civics. Too slow for my

taste.

Negative

I love the look of the current Mustang. It is

certainly reminiscent of the 60s and 70s.

Positive

Bricklin SV1 had a lot of forward-thinking

safety and performance features. Some

ood, some bad.

Mixed

3.2 Data Annotation

In this research, the manual labelling approach was

adopted as it yielded a more elevated degree of

accurate labelling compared to other approaches (Heo

et al., 2021). Three annotators performed the

labelling, and the inter-rater agreement was evaluated

through the Fleiss Kappa index to assess both the

reliability and consistency of the annotators. Fleiss

Kappa is the most widely implemented Kappa in

labelling emotions (Podlesek et al., 2009). The

minimum degree of inter-rater agreement that is

acceptable by many researchers is 80% (McHugh,

2012). In this research, the calculated Fleiss Kappa

degree for the inter-rater agreement was 96.8%,

which denotes almost perfect agreement.

4 PRE-PROCESSING

Pre-processing is the main operation in sentiment

analysis tasks (Awajan et al., 2018). If pre-processing

operations are not performed, it could lead the

analyser to override significant terms. However,

overuse of pre-processing approaches could lead to a

loss of significant data (Mansour et al., 2017). In this

research, the pre-processing operations were divided

into two: normalization, and removal of stop words.

4.1 Normalization

Text extracted from social media is usually not ready

for language processing tasks as it is written using

highly informal language. This informal text needs to

be normalized to an acceptable standard formal style

(Erianda et al., 2017). Hence, the text in this research

was automatically normalized to clean the data,

remove the usernames of the reviewers to ensure

privacy and anonymity, remove punctuation, exclude

non-printable ASCII characters, remove non-English

letters, and convert words from uppercase to

lowercase to reduce the uncertainty that could be

entailed in the classification process. The following

example illustrates the normalization process:

The sentence before normalization: “You bought

a $1200 twenty six year old BMW. Expect plenty of

repairs on an ongoing basis”.

The sentence after normalization: “you bought a

twenty six year old bmw expect plenty of repairs on

an ongoing basis”.

4.2 Removing Stop Words

In many cases, stop words are useless for processing.

Hence, they are discarded to save both size and time

(Awajan et al., 2018). Therefore, a file composed of

around 97 English stop words gathered from (Default

English Stop Words List) was created by us. The

following example illustrates the operation:

The original sentence: “Personally, I think it is

absolutely hideous. My eyes definitely do not see any

elegance with that car”.

The sentence after removing stop words:

“Personally think absolutely hideous eyes definitely

not see elegance car”.

5 FEATURE SELECTION

In this research, the feature selection operations were

conducted to decrease the dimensionality of the

dataset by reducing the initial features and retaining

only important features for classification (Alonso-

Betanzos et al., 2015). Four widely applied feature

selection techniques were utilized, which are: N-

Gram Feature, Lemmatization, POS Tagger (Part-Of-

Speech Tagger), and TFIDF (Term Frequency-

Inverse Document Frequency).

TextBlob Lemmatizer was utilized to lemmatize

the data. It was utilized due to its high accuracy in

producing lemmas, and because it is compatible with

TextBlob POS tagger which was also utilized in this

research to produce POS tags. For illustration, the

following instance demonstrates a lemmatization

operation performed by TextBlob Lemmatizer.

The original sentence: “My wife loves the looking

of Audi A8. She did put her feet on the gas pedal to

enjoy the engine’s sound.”

The sentence post-lemmatization: “My wife love

the look of Audi A8 She do put her foot on the gas

pedal to enjoy the engine’s sound.”

TextBlob was also adopted to carry out the POS

tagging operation as it is fast, easy to use, accessible,

has the highest code quality “L5” (granted by

Lumnify), and holds MIT license (Varma et al.,

2018). Moreover, when we compared three widely

Analysing the Sentiments in Online Reviews with Special Focus on Automobile Market

263

used English POS taggers (spaCy, WordNet, and

TextBlob), the accuracy of identifying and tagging

prepositions, pronouns, and the possessive ending

was the highest with Textblob POS tagger. Table 2

illustrates how a sentence from our dataset is tagged

by TextBlob POS Tagger. The sentence is “I like the

price on that Van.”

Table 2: Example of a Sentence Tagged by TextBlob POS

Tagger.

Word POS Tag Meaning (Penn Treebank II

Set)

I PRP pronoun, personal

Like VBP verb, non-3

person singular

present

The DT determine

Price NN noun, singular, or mass

On IN conjunction, subordinating, or

preposition

That DT determine

Van NNP noun, proper singular

6 SPLITTING AND BALANCING

THE DATASET

The dataset in this research was prepared in three

phases: splitting the dataset between training (70%)

and testing (30%), 10 K-fold cross-validation, and

oversampling the minority classes (the SMOTE-NC

“Synthetic Minority Oversampling Technique

Nominal and Continuous” was used to conduct

synthetic oversampling operation over the classes

with minority occurrence in the training datasets).

7 THE SUPERVISED MACHINE

LEARNING APPROACH

Five machine learning classifiers were used in this

research, which are Linear SVC, Bernoulli Naïve

Bayes, the Multi-layer Perceptron (MLP) Classifier,

CART Decision Tree, and Multinomial Naïve Bayes.

Additionally, two classifiers were also utilized, which

are the Ensemble Vote classifiers (hard and soft) as

shown in Table 3 (Alsafari et al., 2021). The

hyperparameter tuning was conducted to choose the

optimum hyperparameters for the classifiers

rendering the best scores when tested on the training

dataset. Moreover, a study conducted by (Bergstra et

al., 2012) showed that random search theoretically

and empirically resulted in better outcomes when

optimizing hyperparameters as opposed to grid

search. Hence, in this research, the random search

approach was applied to tune the classifiers’

hyperparameters.

Table 3: The F1 Scores of the Machine Learning

Classifiers.

Model Best Parameters and Score F1-

Scor

Linear SVC best_params:{‘max_iter’:

1500, ‘C’: 1}. Best_score:

0.889066

76%

Bernoulli

Naïve Ba

best_params:{‘alpha’:1.0}.best

score: 0.872703

77%

MLP

Classifier

best_params:{‘hidden_layer_si

zes’: 100, ‘alpha’:0.1}.best_

score: 0.908035

77%

CART

Decision Tree

est_params:{‘max_depth’: 14,

‘criterion’: ‘entropy’}.best_

score: 0.785935

68%

Multinomial

Naïve Ba

best_params:{‘alpha’:1.0}.best

score: 0.826234

68%

The Ensemble

Soft Vote

NA 80%

The Ensemble

Hard Vote

NA 78%

Based on the results shown in Table 3, the Ensemble

Soft Vote classifier surpassed the other models by

achieving 80% in F1 scores.

8 THE BERT AND DEEP

LEARNING MODELS

APPROACH

BERT is an advanced and modern word embedding

model developed in 2018 by Google. It was

developed with the purpose of pre-training deep

bidirectional segments from unlabelled data through

combining right and left context in every layer.

Consequently, it is possible to fine-tune a pre-trained

BERT model using a single extra output layer to

generate advanced models for different NLP projects

without making radical architecture adjustments for

each project (Devlin et al., 2019). Therefore, in this

research, BERT was implemented to classify the

sentiments in the dataset (Alsafari et al., 2020).

Moreover, deep learning yields the most optimal

solutions to natural language processing tasks (Prieta

et al., 2020). Therefore, in this paper, the deep

learning approach was implemented in the form of the

ICAART 2022 - 14th International Conference on Agents and Artiﬁcial Intelligence

264

models: CNN, the LSTM, and GRU as shown in

Table 4.

The epoch value was selected to be 50 because it

showed the highest results among the randomly

chosen values. This value was also selected by

(Lehečka et al., 2020) when they classified large-

scale multi-label Wikipedia datasets. Moreover, the

value of the batch size was set to be 32 because it is

an adequate default value according to (Garcia-Silva

et al., 2020). Furthermore, the value of the learning

rate was selected to be 5e-5, which matches the value

selected by (Sun et al., 2019). The stride value was

selected to be 1, which was the same value selected

by (Srivastava et al., 2020), and it is the most popular

value for the stride (Togashi et al., 2018). Moreover,

ReLU (Rectified Linear Unit) was adopted as an

activation function. ReLU was also picked by (Nie et

al. 2020).

The results shown in Table 4 reflect that the

combination of the BERT word embedding model

with the LSTM model surpassed the combination of

the BERT word embedding model with the other deep

learning models by scoring 99.48% in F1-scores. In

general, the combination of the BERT word

embedding model with the deep learning models used

in this research generated quite higher F1 scores than

those achieved by the machine learning classifiers

employed in this research.

Table 4: The F1-Scores Resulted From Combining BERT

Word Embedding Model with a Set of Deep Learning

Models.

Deep Learning

Model

BERT

F1-Score

CNN 99.19%

LSTM 99.48%

GRU 94.73%

BERT 98.22%

9 DATA VISUALIZATION

There is a set of sentiment words and clauses that the

classifiers depend on to determine the polarity of the

classification. For instance, when the Bernoulli Naïve

Bayes classifier analysed the sentiments in the

dataset, the following sets of negative and positive

words and clauses assisted in determining the

sentiment polarity of the text. As stated above, some

of these sentiment words and clauses (such as recall,

of power, and head gasket, among others) are rarely

used outside of the automobile domain. This justifies

limiting the scope of this research to automobile data

as general sentiment analyzers will arguably not be

able to classify these words and clauses. Some of

these words and clauses are illustrated in Figure 1 and

also shown below:

Positive words and clauses:

['more fun', 'world', 'sweet', 'much good', 'it very', 'be

good', 'be great', 'be much', 'white', 'genesis',

'beautiful', 'best', 'best car', 'excellent', 'awesome',

'nice', 'overall', 'look great', 'comfy', 'one of most', 'one

of best', 'very nice', 'car but', 'safe', 'very comfortable',

'of power', 'fantastic', 'fine', 'of best', 'fun drive'].

Negative words and clauses:

['car but', 'terrible', 'not like', 'crap', 'break',

'unreliable', 'crappy', 'fall', 'noise', 'failure',

'uncomfortable', 'awful', 'ugly', 'certain', 'recall', 'po',

'pricey', 'fail', 'but be', 'and not', 'more expensive', 'but

not', 'gasket', 'horrible', 'head gasket', 'worst', 'hate',

'hat', 'weak', 'poor'].

10 CONCLUSION AND FUTURE

WORK

Despite the improvements suggested for future work

outlined below, it is evident that the methodology

used in this research successfully filled the gaps that

were left unaddressed by other contributions

regarding the analysis of sentiments that exist in

reviews in the automobile domain. In terms of the

results, the combination of the BERT word

embedding model with the LSTM model had the

highest F1-score, which reflects an opportunity for

researchers to adopt such a combination to analyse

the sentiments in English automobile data in

particular, and non-English automobile data in

general. The methodology adopted in this research

has also shown superior F1 scores when compared

with the scores achieved by other works that were

reviewed in this paper.

In future work, adopting advanced models such as

reinforcement learning, ERNIE, and Elmo could

enhance the results and widen the scope of this area

of research. The results could also be improved by

enlarging the dataset, treating negations, and

developing specific word embedding models that are

more related to the automobile industry in terms of

the embedded vocabulary. The scope of the work

could also be broadened by covering the semi-

supervised approaches (Alsafari et al., 2021). Finally,

performing an aspect-based sentiment analysis would

result in more precise sentiment analysis for the

opinion target.

Analysing the Sentiments in Online Reviews with Special Focus on Automobile Market

265

Figure 1: The Negative and Positive Words and Clauses that Assisted the Classifier in Determining the Sentiment Polarity of

the Text.

REFERENCES

Wijnhoven, F., Plant, O. (2017). Sentiment Analysis and

Google Trends Data for Predicting Car Sales. The

Thirty Eighth International Conference on

Information Systems.

Autos.ca. [cited 05-01-2021]; Available from:

https://www.autos.ca/forum/index.php?board=6 7.0

English Automobiles Dataset. [cited 24-9-2021]; Available

from: https://github.com/aymanya/English-Automobi

les-Dataset

Yafoz, A., Mouhoub, M. (2021). Sentiment Analysis in

Arabic Social Media Using Deep Learning Models.

In Proceedings of the IEEE International Conference

on Systems, Man, and Cybernetics (IEEE SMC 2021).

Malik, K. Khan, M. (2018). Sentiment Classification of

Customer’s Reviews about Automobiles in Roman

Urdu. In Proceedings of the 2018

Future of

Information and

Communication

Conference (FICC).

Alsawalqah, H., Aljarah, I., Yaghi, R., Shukri, S. (2015).

Twitter Sentiment Analysis: A Case Study in the

Automotive Industry. 2015 IEEE Jordan Conference

on Applied Electrical Engineering and Computing

Technologies (AEECT).

Yafoz, A., Mouhoub, M. (2020). Analyzing

MachineLearning Algorithms for Sentiments in

ArabicText. In Proceedings of the IEEE International

Conference on Systems, Man, and Cybernetics (IEEE

SMC 2020).

Octoparse Web Scraper. [cited 24-02-2021]; Available

from https://www.octoparse.com/

Heo, G., Whang, S., Roh, Y. (2021). A Survey on Data

Collection for Machine Learning: A Big Data - AI

Integration Perspective. IEEE Transactions on

Knowledge and Data Engineering Journal. 33(4).

Podlesek, A., Boštjan, V., Mihelič, F., Komidar, L., Gajšek,

R., Štruc, V. (2009). Analysis and Assessment of AvID:

Multi-Modal Emotional Database. In International

Conference on Text, Speech and Dialogue (TSD 2009).

McHugh, L. (2012). Interrater Reliability: The Kappa

Statistic. The Croatian Society of Medical Biochemistry

and Laboratory Medicine (Biochemia Medica) Journal.

22(3).

Awajan, I., Mohamad, M. (2018). A Review on Sentiment

Analysis in Arabic Using Document Level.

International Journal of Engineering and Technology.

7((3.13) (2018)).

Mansour, H., El-Masri, M., Altrabsheh, N. (2017).

Successes and Challenges of Arabic Sentiment

Analysis Research: A Literature Review. Social

Network Analysis and Mining Journal. 7(1).

Erianda, A., Rahmayuni, I. (2017). Improvement of Email

and Twitter Classification Accuracy Based on

Preprocessing Bayes Naïve Classifier Optimization

in Integrated Digital Assistant. The International

Journal on Informatics Visualization. 1(2).

Default English Stop Words List. [cited 03-10-2021];

Available from: https://www.ranks.nl/stopwords

Alonso-Betanzos, A., Sánchez-Maroño, N., Bolón-Canedo,

V. (2015). Feature Selection for High Dimensional

Data (Artificial Intelligence: Foundations, Theory, and

Algorithms). Springer. 1st Edition.

Varma, A., Saha, A., Kunal, S., Tiwari, V. (2018). Textual

Dissection of Live Twitter Reviews Using Naïve

Bayes. Procedia Computer Science Journal.

132(2018).

ICAART 2022 - 14th International Conference on Agents and Artiﬁcial Intelligence

266

Penn Treebank II Tag Set. [cited 07-04-2021]; Available

from: http://relearn.be/2015/training-common-sense/

sources/software/pattern-2.6-critical-

fork/docs/html/mbsp-tags.html

Alsafari, S., Sadaoui, S. (2021). Ensemble-based Semi-

Supervised Learning for Hate Speech Detection.

Proceedings of the Thirty-Fourth International Florida

Artificial Intelligence Research Society Conference.

Bergstra, J., Bengio, Y. (2012). Random Search for Hyper-

Parameter Optimization. Journal of Machine Learning

Research. 13(2012).

Devlin, J., Lee, K., Toutanova, K., Chang, M. (2019).

BERT: Pre-training of Deep Bidirectional

Transformers for Language Understanding. In

Proceedings of the Annual Conference of the North

American Chapter of the Association for

Computational Linguistics (NAACL-HLT).

Alsafari, S., Sadaoui, S., Mouhoub, M. (2020). Hate and

Offensive Speech Detection on Arabic Social Media.

Online Social Networks and Media Journal. 19.

Prieta, F., Moreno-García, M., Dang, N. (2020). Sentiment

Analysis Based on Deep Learning: A Comparative

Study. Electronics Journal. 9(3,483).

Lehečka, J., Švec, J., Šmídl, L., Ircing, P. (2020). Adjusting

BERT’s Pooling Layer for Large-Scale Multi-Label

Text Classification. International Conference on Text,

Speech, and Dialogue (TSD 2020).

Garcia-Silva, A., Gomez-Perez, J., Denaux, R. (2020). A

Practical Guide to Hybrid Natural Language

Processing: Combining Neural Models and Knowledge

Graphs for NLP. Springer.

Sun, C., Qiu, X., Xu, Y., Huang, X. (2019). How to Fine-

Tune BERT for Text Classification. China National

Conference on Chinese Computational Linguistics

(CCL 2019).

Srivastava, H., Srivastava, S., Kumari, S., Varshney, V.

(2020). A Novel Hierarchical BERT Architecture for

Sarcasm Detection. Proceedings of the Second

Workshop on Figurative Language Processing.

Togashi, K., Nishio, M., Do, R., Yamashita, R. (2018).

Convolutional Neural Networks: An Overview and

Application in Radiology. Insights into Imaging

Journal. 9(4).

Nie, J., Du, P., Lu, Z. (2020). VGCN-BERT: Augmenting

BERT with Graph Embedding for Text Classification.

European Conference on Information Retrieval (ECIR

2020).

Alsafari, S., Sadaoui, S., Mouhoub, M. (2021). Semi-

Supervised Self-Training of Hate and Offensive Speech

from Social Media. Applied Artificial Intelligence

Journal.

Analysing the Sentiments in Online Reviews with Special Focus on Automobile Market

267