AMAZON ECHO REVIEWS SENTIMENT ANALYSIS

USING NLP
This project report is submitted in partial fulfilment of the requirement

for the award of the degree

BACHELOR OF TECHNOLOGY
IN
COMPUTER SCIENCE AND ENGINEERING

Submitted by
JILAGAM DHEERAJ
Reg.no: 208297601021
Under the esteemed guidance of
Dr B. Kezia Rani
Associate Professor & Head of the Department
DEPARTMENT OF COMPUTER SCIENCE AND ENGINEERING
ADIKAVI NANNAYA UNIVERSITY
RAJAMAHENDRAVARAM
2020-2024
 ADIKAVI NANNAYA UNIVERSITY
RAJAMAHENDRAVARAM
UNIVERSITY COLLEGE OF ENGINEERING
DEPARTMENT OF COMPUTER SCIENCE & ENGINEERING

CERTIFICATE
This is to certify that this project report entitled, “AMAZON ECHO REVIEWS
SENTIMENT ANALYSIS USING NLP” is a bonafide work of JILAGAM
DHEERAJ, Reg.No: 208297601021 submitted in a partial fulfilment of the
requirements for the award of Degree of BTech (CSE) during the period 2020-2024.
This work carried out by him under my supervision and guidance and submitted to
Department of Computer Science and Engineering, Adikavi Nannaya University.

INTERNAL GUIDE HEAD OF THE DEPARTMENT

EXTERNAL EXAMINER
 DECLARATION

I JILAGAM DHEERAJ, reg.no:208297601021, hereby declare that the project report

entitled “Amazon Echo Reviews Sentiment Analysis Using NLP” done by me under
the guidance of Dr B.Kezia Rani, Associate Professor, Adikavi Nannaya University,
is submitted for the partial fulfilment of requirement of the award of the degree,
Bachelor of Technology in Computer Science and Engineering in the academic year
2020-2024.

Signature of the student

 ACKNOWLEDGEMENT

I feel fortunate to pursue my Bachelor of Technologies degree in the Department of

Computer Science and Engineering, University College of Engineering, Adikavi
Nannaya University. It provided all the facilities in the area of Computer Science and
Engineering. It’s a genuine pleasure to express my deep sense of thanks and gratitude
to my mentor and project guide Dr B.Kezia Rani, Associate Professor, Department
of Computer Science and Engineering, Adikavi Nannaya University for her
excellent guidance right from the selection of the project and her valuable suggestions
throughout the project work. Her constant and timely counsel has caused me to succeed
in completing this project in college.
I profusely thank Dr. B. Kezia Rani, Head of the Department, of Computer Science
and Engineering for the guidance.
I also thank Dr. P. Venkateswara Rao, Principal, University College of Engineering
for all the encouragement and support.
I also thank Dr. D. Latha, Project Coordinator, Department, Computer Science and
Engineering for the guidance.
A great deal of thanks goes to Review Committee Members and the entire Faculty of
the Department of Computer Science and Engineering, University College of
Engineering, Adikavi Nannaya University, Rajamahendravaram for their excellent
supervision and valuable suggestions.

JILAGAM DHEERAJ
208297601021
 ABSTRACT
The increasing volume of product reviews across online platforms presents a challenge
for sellers to analyze customer sentiments. Examining thousands of customer reviews
is time-consuming and tedious. This project proposes a sentiment analysis system
utilizing machine learning (ML) and natural language processing (NLP) to address the
challenge that benefits both the company and the customer. The system aims to extract
valuable insights from large sets of product reviews, aiding sellers in enhancing product
features and informing potential customers about product pros and cons. This project
enhances the present scenario by offering NLP techniques to train ML models using
XG Boost and Random Forest models, and advanced data visualizations providing
insights to customers. Effective customer feedback analysis is crucial for any company,
as it serves as a valuable resource for enhancing business opportunities and overall
improvement.
Key-words: XG Boost, Random Forest, Sentiment Analysis, NLP
 INDEX

PAGE
CHAPTER TITLE
NO
ABSTRACT
1. INTRODUCTION 01
1.1. Existing System 02
1.2. Proposed System 02
2. SYSTEM ANALYSIS
2.1. Functional Requirements 03
2.2 Non-Functional Requirements 03
2.3 Software Requirements 04
2.4 Hardware Requirements 04
3. SYSTEM DESIGN
3.1 System Architecture 05
3.2 UML Diagrams 07
3.2.1 Use case Diagram 07
3.2.2 Class Diagram 09
3.2.3 Sequence Diagram 10
3.2.4 Activity Diagram 11
3.2.5 Deployment Diagram 12
4. SYSTEM IMPLEMENTATION
4.1. Data Set 13
4.2. Statistical Measures 14
4.3. Data Preprocessing and NLP Techniques 15
4.4. Code and Implementation 16
4.5. Classification Reports 18
4.6. Testing 20
5. OUTPUT SCREENS 21
6. CONCLUSION 24
7. FUTURE SCOPE 25
8. LITERATURE SURVEY 26
9. BIBLIOGRAPHY 30
 LIST OF FIGURES

Fig.No DESCRIPTION PAGE

3.1 System Architecture 05

3.2.1 Use case Diagram 07

3.2.2 Class Diagram 09

3.2.3 Sequence Diagram 10

3.2.4 Activity Diagram 11

3.2.5 Deployment Diagram 12

4.1.1 Data Selection 13

4.2.1 Statistical Measures 14

4.2.2 Positive and Negative Length Graph 15

5.1 Flask Interface 21

5.2 Positive Sentiment Example 21

5.3 Negative Sentiment Example 22

5.4 Bulk Sentiment Example 22

5.5 Graph Generation 23

5.6 Word Cloud Generation 23

0
 1. INTRODUCTION
The project includes several key components, such as Data Selection, Data
Preprocessing, NLP techniques, Data Splitting, Feature Scaling, Classification, and
Result Generation. NLP techniques are used to preprocess the text data. This involves
tasks such as tokenization, lowercasing, removing stop words, stemming, and handling
special characters. The preprocessed data is split into training and testing sets. Machine
learning models, such as XGBoost and Random forest classifiers, are applied to the
preprocessed and split data to perform sentiment analysis. The performance of the
applied models is evaluated using metrics such as accuracy, precision, recall, and F1
score to assess their effectiveness in predicting sentiment.

The dataset is split into training and testing sets using the train_test_split function,
where 70% of the data is used for training and 30% of the data is used for testing. Model
Training as the sentiment analysis model is trained using different classifiers, such as
XGBoost, Random Forest, and Decision Tree Classifier. Model Evaluation is the
accuracy of the models on both the training and testing datasets to assess their
performance. Saving Models and Transformers as the trained models Random Forest,
XGBoost, and preprocessing transformers CountVectorizer, and MinMaxScaler are
saved using the pickle module. This allows the use of these models later without
retraining.

The project also involves the development of a user interface, potentially using a web-
based application built with Flask, to allow users to interact with the sentiment analysis
system. Web Application Integration for user interaction, an interface using Flask is
created, allowing users to upload a CSV file for bulk prediction or enter text for single
predictions. The interface supports functionalities such as selecting the dataset,
providing text input for analysis, and displaying the sentiment analysis results.

1
1.1 EXISTING SYSTEM

The existing system is a rule-based system that uses pre-defined rules to classify the
sentiment of the reviews. This approach involves defining a set of rules based on
keywords or phrases that indicate the sentiment. However, this approach may not be
effective in capturing the nuances of language and may not be adaptable these days.

DRAWBACKS:

Nuance and Context issues

Low Scalability and High Maintenance

Limited Adaptability

1.2 PROPOSED SYSTEM

By adopting machine learning and NLP-based approaches, the project aimed to

overcome the limitations of the rule-based approach, allowing for more adaptive, data-
driven, and context-aware sentiment analysis of the Amazon Alexa reviews. The
interface supports functionalities such as selecting the dataset, providing text input for
analysis, and displaying the sentiment analysis results.

ADVANTAGES:

Better Nuance and Context

Efficient for larger data sets

Subjective

2
 2. SYSTEM ANALYSIS

2.1 FUNCTIONAL REQUIREMENTS

1) Prediction: The sentiment analysis system must be able to utilize the trained model
to make accurate predictions on new data, ensuring that it can consistently provide
reliable and up-to-date results.

2) Graph generation: It is essential for the system to generate graphs and reports that
summarize sentiment trends and patterns in a visually appealing and user-friendly way.
This feature enables users to interpret the data quickly and make informed decisions
based on the results.

3) Data acquisition: The system should be designed to allow users to interact with it
and input their own sentences for sentiment analysis. This feature will enable the system
to gather more data and improve its performance.

4) Evaluation Metrics: To ensure that the sentiment analysis model performs

accurately, the system must use standard evaluation metrics such as precision, recall,
and F1 score to evaluate its performance continually. This feature ensures that the
system remains reliable and provides accurate results.

2.2 NON-FUNCTIONAL REQUIREMENTS

1) Usability: The user interface must be intuitive, with clear navigation and easy-to-
understand features to ensure user adoption and satisfaction.

2) Reliability: The code should handle potential errors gracefully, providing

informative error messages.

3) Cross-Platform Compatibility: The code should be portable across different

operating systems like Windows, macOS, and Linux without significant modifications.

4) Ease of Understanding: Code and comments should be in a way that is easy to

understand for both novice and experienced users.

3
2.3 SOFTWARE REQUIREMENTS

1)Programming Language: Python-3

2)Integrated Development Environment (IDE): Jupyter Notebook, Pycharm.

3)Libraries and Packages: pandas for data manipulation

wordcloud for creating word clouds

scikit-learn for machine learning tools

matplotlib and seaborn for data visualization

JupyterThemes for distinguishing between graphs

nltk for natural language processing tasks

XGBoost For implementing the XGBoost classifier for

sentiment analysis.

2.4 HARDWARE REQUIREMENTS

1)RAM: 4GB

2)Storage: 32 GB

3)Processor: Intel i5

4
 3. SYSTEM DESIGN

3.1 SYSTEM ARCHITECTURE

FIG 3.1 SYSTEM ARCHITECTURE

1. Data Acquisition: The process starts with acquiring data from an Amazon
review dataset.
2. Preprocessing: The data is then preprocessed to prepare it for modeling. This
includes:
o Dropping unwanted columns: Irrelevant columns are removed from the
dataset.
o Text cleaning: This may involve removing punctuation, stop words, and
stemming or lemmatization (converting words to their base form).
3. Feature extraction: Features are extracted from the text data. This could
involve techniques like:

5
 o Bag-of-words: This converts each document into a sparse vector where
each element represents the number of times a particular word appears
in the document.
o TF-IDF: This is a variant of bag-of-words that weights the importance
of words based on their frequency in a document and rarity across the
corpus.
4. Model training: The extracted features are then used to train a machine learning
model to predict sentiment. The diagram shows two models being trained:
XGBoost and Random Forest. These are both popular tree-based ensemble
methods for classification.
5. Testing and evaluation: The evaluation process involves feeding the test set
into the model and measuring its performance metrics, such as accuracy,
precision, recall, and F1 score. The results of the evaluation will help determine
whether the model is ready for deployment or if further improvements are
needed.
6. Model Predictions: Once a machine learning model is trained and evaluated, it
can be used to generate predictions on new and unseen data. This process
involves applying the model to the new data, which could be in the form of text,
images, or any other input format that the model was designed to handle. For
example, if the model was trained on Amazon reviews and labeled them as
positive or negative based on their sentiment, it can be used to predict the
sentiment of new Amazon reviews. This way, the model can assist in automating
tasks that would otherwise require human effort.

7. Visualizations: After conducting an in-depth analysis of the data, you can use
various visualization techniques to represent the findings in a more meaningful
way. These visualizations can provide a clear picture of the sentiment of the data
and help you gain valuable insights into the patterns and trends that might have
been missed otherwise. By using charts, graphs, and other visual tools, you can
present the results in a more compelling and easy-to-understand manner,
enabling you to make better-informed decisions based on the data analysis.

6
 3.2 UML DIAGRAMS

3.2.1 CLASS DIAGRAM

The class diagram is a type of static structure diagram that describes the structure of a
system by showing the system's classes, their attributes, operations or methods, and the
relationships among objects.

Fig:3.2. 1 CLASS DIAGRAM

1. Input: This class is responsible for handling user input, including selecting the data
to be analyzed and providing text input for sentiment analysis. Select data(): This
method allows the user to select the dataset to be analyzed for sentiment analysis.
Provide_text(): This method allows the user to input text for sentiment analysis.

7
2. Preprocessing: This class is responsible for preprocessing the input data, including
handling missing values and vectorizing the data for machine learning model
application. Missing_values(): This method handles missing values in the input data,
potentially using techniques such as imputation or deletion. Vectorization(): This
method vectorizes the input data, converting it into a format suitable for machine
learning model applications.
3. Data Splitting: This class is responsible for splitting the preprocessed data into
training and testing sets, a crucial step in machine learning model development. Test():
This method splits the preprocessed data into a testing set, which is used to evaluate the
performance of the machine learning model. Train(): This method splits the
preprocessed data into a training set, which is used to train the machine learning model.
4. Classification: This class is responsible for applying machine learning models to the
preprocessed and split data for sentiment analysis. XGBoost(): This method applies the
XGBoost machine learning model to the preprocessed and split data for sentiment
analysis. Random Forest(): This method applies the Random Forest machine learning
model to the preprocessed and split data for sentiment analysis.
5. Performance Analysis: This class is responsible for evaluating the performance of
the applied machine learning models and generating sentiment analysis predictions.
Prediction(): This method generates sentiment analysis predictions based on the applied
machine learning models, potentially using metrics such as accuracy, precision, recall,
or F1 score to evaluate performance.

3.2.2 USE CASE DIAGRAM

A use case diagram is a visual representation of how a user interacts with a system. It
shows the various use cases and user types for a system. Use case diagrams are a type
of behavior diagram.
1. User: The primary actor in the system, interacting with the client interface to provide
input for sentiment analysis. The user initiates the sentiment analysis process by
providing input data and expects to receive the sentiment analysis results.
2. System: The system itself, comprises various components responsible for processing
the user input, performing sentiment analysis, and presenting the results back to the
user.
3. Dataset: The system acquires the dataset containing Amazon Alexa reviews, which
serves as the basis for sentiment analysis. Preconditions: The dataset must be available
for the sentiment analysis process to begin. Postconditions: The dataset is ready for
further processing and analysis.
4. Data Preprocessing: The system preprocesses the input data, which involves tasks
such as removing special characters, converting text to lowercase, and tokenizing the
input into individual words. Preconditions: The dataset must be available, and the
preprocessing components must be operational. Postconditions: The preprocessed data
is ready for further analysis and model application.

8
5. Split the Data: The system splits the preprocessed data into training and testing sets,
a crucial step in machine learning model development. Preconditions: The preprocessed
data must be available and ready for splitting. Postconditions: The training and testing
sets are prepared for model training and evaluation.

FIG 3.2.2 USE CASE DIAGRAM

6. Apply Model: The system applies the machine learning model like XGBoost or
Random Forest classifier to the preprocessed and split data for sentiment analysis.
Preconditions: The preprocessed and split data must be available, and the machine
learning model must be trained and ready for application. Postconditions: The sentiment
analysis results are generated based on the applied model.
7. Evaluation: The system evaluates the performance of the applied model, typically
using metrics such as accuracy, precision, recall, or F1 score. Preconditions: The
sentiment analysis results and ground truth labels must be available for evaluation.
Postconditions: The model's performance metrics are determined, providing insights
into its effectiveness.
8. Display Results: The system displays the sentiment analysis results to the user,
potentially including visualizations such as word clouds or graphs to represent the
sentiment distribution. Preconditions: The sentiment analysis results must be available
and ready for presentation. Postconditions: The sentiment analysis results are presented
to the user for review and interpretation.

9
3.2.3 SEQUENCE DIAGRAM
A sequence diagram illustrates the sequence of messages between objects in an
interaction. A sequence diagram consists of a group of objects that are represented
by lifelines and the messages that they exchange over time during the interaction.

FIG 3.2.3 SEQUENCE DIAGRAM

1. User Input Interaction: The sequence diagram begins with the user interacting with
the client interface, providing input in the form of text or uploading a CSV file for
sentiment analysis.
2. Client-Server Interaction: Upon receiving the user input, the client interface sends
a request to the server, either as a text input or a file upload request.
3. Server-Side Processing: The server receives the user input request and initiates the
sentiment analysis process. This involves pre-processing the input data, transforming it
using the CountVectorizer and MinMaxScaler, and passing it to the machine learning
model for sentiment prediction.
4. Machine Learning Model Interaction: The server interacts with the machine
learning model to perform sentiment analysis on the pre-processed input data. This
interaction includes passing the transformed data to the model and receiving the
predicted sentiment results.
5. Result Visualization and Output: After sentiment analysis, the server may generate
visualizations and prepare the sentiment analysis results for presentation to the user.
6. Client Response: The server sends the sentiment analysis results back to the client
interface, which then displays the results to the user, either as individual sentiment
predictions or as a downloadable CSV file.

10
3.2.4 ACTIVITY DIAGRAM
An activity diagram shows the flow from one activity to another in a system or process.
It's used to describe the different dynamic aspects of a system and is referred to as a
'behavior diagram' because it describes what should happen in the modeled system.

FIG:3.2. 4 ACTIVITY DIAGRAM

1. User Input Processing: The activity diagram can start with a user providing input to
the system, either as a text input or by uploading a CSV file containing multiple
sentences for sentiment analysis.
2. Input Validation and Pre-processing: Upon receiving the user input, the system
validates the input to ensure it meets the required format and data integrity. For text
input, the system may perform pre-processing steps such as removing special
characters, converting text to lowercase, and tokenizing the input into individual words.
- If a CSV file is uploaded, the system reads and validates the file, extracting the relevant
data for sentiment analysis.
3. Sentiment Analysis: The pre-processed input data is then passed through the
sentiment analysis pipeline, which involves applying the CountVectorizer and
MinMaxScaler to transform the input data into a format suitable for the machine
learning model. - The transformed data is then fed into the trained machine learning

11
model like, XGBoost or Random Forest classifier to predict the sentiment of each input
sentence.
4. Result Visualization and Output: After sentiment analysis, the system may
generate visualizations such as word clouds or graphs to represent the sentiment
distribution of the input data. - The system then provides the sentiment analysis results
to the user, either by displaying the predicted sentiment for each input sentence or by
generating a downloadable CSV file containing the predictions.

3.2.5 DEPLOYMENT DIAGRAM

FIG:3.2 5 DEPLOYMENT DIAGRAM

NODES:
Browser: Represents the user's web browser.
Server: Represents the local machine where the Flask app is deployed.
Flask App: Represents the Flask application itself, containing its components.
ARTIFACTS:
app.py: The main Flask application script.
run.py: The script used to start the Flask development server.
Machine Learning Model: The trained machine learning model used by the app.
Dataset: The dataset used to train the machine learning model.

12
 4. SYSTEM IMPLEMENTATION

4.1 DATASET:

The dataset for this project is sourced from Amazon’s official Alexa skill store. The
dataset encompasses the 5th generation of Echo Dot reviews, ranging from 2023, It
contains a total of 3150 customer reviews. The dataset is in .tsv format and can be
processed using the pandas library.

Useful Attributes: Verified_Reviews, Feedback, Length.

Useless Attributes: Date, Variation.

Fig 4.1.1 Data set selection

13
4.2 STATISTICAL MEASURES:

Fig 4.2.1 Statistical Measures

Count: Total number of non-null values in the column. In this case, there
are 3150 non-null values.

Mean: The average value of the column. The average length is

approximately 132.67.

Std: The standard deviation, which measures the dispersion or spread of

the values. It's approximately 182.53, indicating a

wide variation in the lengths.

Min: The minimum value in the column. The shortest length is 1.

25%: The value below which 25% of the data falls. Also known as the
first quartile or Q1. In this case, 25% of the data have a length of 30 or
less.

50%: The median of the data, also known as the second quartile or Q2.
Half of the data have a length of 74 or less.

14
75%: The value below which 75% of the data falls. Also known as the
third quartile or Q3. In this case, 75% of the data have a length of 166 or
less.

Max: The maximum value in the column. The longest length is 2853.

Fig 4.2.2 Positive and Negative Reviews Length Graph

4.3 DATA PREPROCESSING AND NLP TECHNIQUES:

1. Stopword Removal: This involves removing common words that do not

carry much meaning, such as "the", "and", "a", etc. In this project, the
`nltk.corpus.stopwords` module is used to remove stopwords from the text
data.

2. Stemming: This involves reducing words to their root form, such as

"running" to "run". In this project, the `nltk.stem.PorterStemmer` module
is used to stem the words in the text data.

3. Text Cleaning: This involves removing any unwanted characters,

symbols, or punctuation from the text data. In this project, the `re.sub()`
method is used to replace any non-alphabetic characters with a space.

15
4. Feature Scaling: This involves scaling the numerical features in the data
to a common scale to avoid bias towards certain features. In this
project,`sklearn.preprocessing.MinMaxScaler` module is used to scale the
numerical features in the data.

4.4 CODE

a) Data Preprocessing and NLP techniques code:

b) Data Splitting and Code:

Splitting the Data: The dataset is split into training and testing sets using
the train_test_split function, where 70% of the data is used for training
and 30% of the data is used for testing.

16
c) Implementation:

17
4.5 CLASSIFICATION REPORTS:

a) XGBOOST Classification Report

b) RANDOM FOREST Classification Report

c) DECISION TREE Classification Report

18
 5. OUTPUT SCREENS

Fig 5.1 Flask Interface

Fig 5.2 Positive Sentiment Example

19
Fig 5.3 Negative Review Example

Fig 5.4 Bulk Sentiment Prediction

20
 Fig 5.5 Graph Generation

Fig 5.6 Word Cloud Generation

21
 6. CONCLUSION

The model aimed to enhance the sentiment analysis of Amazon Echo reviews by
utilizing machine learning and natural language processing (NLP) techniques. By
adopting this approach, a more adaptable, data-driven, and context-aware methodology
was incorporated, which effectively overcame the limitations of a rule-based approach.
Additionally, the interface offers several functionalities such as dataset selection, text
input, and sentiment analysis result display.

The project comprises several critical components such as Data Selection, Data
Preprocessing, NLP techniques, Data Splitting, Feature Scaling, Classification, and
Result Generation. NLP techniques are used to preprocess the text data by executing
tasks such as tokenization, lowercasing, removing stop words, stemming, and handling
special characters. The preprocessed data is then split into training and testing sets.
Machine learning models including XGBoost and Random forest classifiers are applied
to the preprocessed and split data to perform sentiment analysis. The efficiency of these
models is evaluated using metrics such as accuracy, precision, recall, and F1 score to
determine their effectiveness in predicting sentiment.

22
 7. FUTURE SCOPE

In the future, it is possible to incorporate more advanced Natural Language Processing

(NLP) libraries such as spaCy into sentiment analysis operations. By incorporating pre-
trained language models, the feature extraction and sentiment analysis processes can be
improved. This would enable the implementation of real-time sentiment analysis
capabilities to analyze sentiments from live Amazon reviews as they are posted,
providing timely insights to users. Moreover, the use of aspect-level sentiment analysis
would allow for a more in-depth understanding of specific aspects or features
mentioned in the reviews, like product quality and customer service. This level of detail
could provide essential insights into what customers like or dislike, enabling businesses
to make more informed decisions about their products and services These libraries offer
pre-trained language models that could more effectively capture the nuances of human
language, resulting in more accurate sentiment analysis.

Furthermore, the integration of more advanced NLP libraries into sentiment analysis
operations could revolutionize the way businesses analyze customer feedback. By
providing more accurate and detailed insights, businesses could improve their products
and services, resulting in happier customers and increased revenue.

23
 8. BIBLIOGRAPHY:

[1] A. Bandi and A. Fellah, “Socio-analyzer: A sentiment analysis using social media
data,” in Proc. 28th Int. Conf. Softw. Eng. Data Eng., in EPiC Series in Computing, vol.
64, F. Harris, S. Dascalu, S. Sharma, and R. Wu, Eds. Amsterdam, The Netherlands:
EasyChair, 2019, pp. 61–67.

[2] U. Naseem, S. K. Khan, M. Farasat, and F. Ali Abusive language detection:

A comprehensive review. Boca Raton, FL, USA: CRC Press, 2013.

[3] Krishnendu Chakraborty, Machine Learning with Real World Projects. Infosys
wingspan.

[4] C. C. Aggarwal and C. K. Reddy, Data Clustering: Algorithms and Applications.

Boca Raton, FL, USA: CRC Press, 2013.

[5] N. Ahmad and J. Siddique, “Personality assessment using Twitter tweets,” Procedia
Comput. Sci., vol. 112, pp. 1964–1973, Sep. 2017.

[6] T. Ahmad, A. Ramsay, and H. Ahmed, “Detecting emotions in English and Arabic
tweets,” Information, vol. 10, no. 3, p. 98, Mar. 2019.

24