Download as docx, pdf, or txt
Download as docx, pdf, or txt
You are on page 1of 88

AN APPROACH TO ENHANCING COURSE

EVALUATION WITH ASPECT-BASED


SENTIMENT ANALYSIS: REVEALING
MULTIFACETED STUDENT FEEDBACK
By
ADEKOLU SAMUEL OLUWASEUN
VUG/CSC/20/4035

A PROJECT SUBMITTED TO

THE DEPARTMENT OF COMPUTER AND INFORMATION TECHNOLOGY


FACULTY OF NATURAL AND APPLIED SCIENCES, VERITAS UNIVERSITY,
ABUJA

IN PARTIAL FULFILMENT OF THE REQUIREMENTS FOR THE AWARD OF THE


BACHELOR OF SCIENCE DEGREE (BSc) IN COMPUTER SCIENCE

23RD May, 2024


CERTIFICATION

This is to certify that this project entitled “AN APPROACH TO ENHANCING COURSE

EVALUATION WITH ASPECT-BASED SENTIMENT ANALYSIS: REVEALING

MULTIFACETED STUDENT FEEDBACK” was carried out by “ADEKOLU SAMUEL

OLUWASEUN“ with matriculation number “VUG/CSC/20/4035” in the Faculty of Natural and

Applied Sciences, Veritas University, Abuja for the award of Bachelor of Science in Computer Science.

Mr Felix Uloko __________________

(Project Supervisor) Sign and Date

Dr. Joe Essien __________________

Sign and Date


(Head of Department)

Title Page

Prof. Simon Kowa Okwute __________________

Sign and Date


(Dean of Faculty)

__________________

PROF. OLUMIDE OWOLABI (EXTERNAL Sign and Date


EXAMINER)

II
DECLARATION

This is to declare that this research project titled “AN APPROACH TO ENHANCING
COURSE EVALUATION WITH ASPECT-BASED SENTIMENT ANALYSIS:
REVEALING MULTIFACETED STUDENT FEEDBACK”, carried out by “ADEKOLU
SAMUEL OLUWASEUN VUG/CSC/20/4035”, is solely the result of my work except
otherwise stated.

___________________________________________

ADEKOLU SAMUEL OLUWASEUN, VUG/CSC/20/4035

___________________________________________

DATE

III
DEDICATION

I dedicate this work first of all to God Almighty, the fountain and source of my inspiration and
knowledge, for guidance and protection over me. I also dedicate this work to my lovely parents,
MR AND MRS ADEKOLU, for their financial support and fervent prayers, which have
sustained me throughout this journey, your unwavering support has motivated me to give my all
to this endeavor.

IV
ACKNOWLEDGEMENT

First and foremost, I express my gratitude to the almighty God for the provision of life, wisdom,
good health, and strength throughout this period. I am deeply indebted to my project supervisor,
Mr. Felix Uloko, for his invaluable suggestions, guidance, and unwavering support. His expertise
and encouragement have been instrumental in shaping this project. I would also like to
acknowledge the Head of the Department of Computer and Information Technology, Dr. Joe
Essien, for his outstanding leadership, support, and kindness throughout my academic journey.
His dedication to excellence has inspired me greatly. Additionally, I extend my appreciation to
all the staff and lecturers of Programming for their assistance and guidance over the years.

Finally, I am grateful to my parents for their prayers, relentless love, kindness, and
understanding. Their unwavering support has been a constant source of strength and motivation.

To Mr. Felix Uloko, Dr. Joe Essien, and all those mentioned above, I say a heartfelt thank you.

V
ABSTRACT

This research shows the integration of aspect-based sentiment analysis to enhance the process of
course evaluation by uncovering nuanced student feedback. Traditional course evaluations often
provide limited insights into various aspects of teaching and learning. In contrast, this study
proposes a methodology that employs aspect-based sentiment analysis to dissect student
feedback into distinct facets/aspects, enabling a comprehensive understanding of their sentiments
towards different aspects of a Course. By categorizing feedback into specific dimensions such as
teaching style, content relevance, assessment methods, and more, educators can gain valuable
insights into strengths and areas for improvement. This approach not only offers a more detailed
perspective on student opinions but also facilitates data-driven decision-making for curriculum
enhancement and instructional refinement. Through the incorporation of aspect-based sentiment
analysis, this research contributes to a more holistic and effective course evaluation process.

VI
TABLE OF CONTENTS

TITLE PAGE --------------------------------------------------------------------------------------------- I

CERTIFICATION ---------------------------------------------------------------------------------------II
DECLARATION ----------------------------------------------------------------------------------------III
DEDICATION-------------------------------------------------------------------------------------------- IV
ACKNOWLEDGEMENT ----------------------------------------------------------------------------- V
ABSTRACT --------------------------------------------------------------------------------------------- VI
TABLE OF CONTENTS ---------------------------------------------------------------------------- VII
LIST OF FIGURES ---------------------------------------------------------------------------------- XI
LIST OF TABLES ------------------------------------------------------------------------------------ XII
LIST OF APPENDIX ---------------------------------------------------------------------------------XIII

CHAPTER ONE............................................................................................................................1

INTRODUCTION.........................................................................................................................1

1.1. BACKGROUND OF THE STUDY...................................................................................1

1.2. STATEMENT OF THE PROBLEM.................................................................................1

1.3. RESEARCH QUESTIONS.................................................................................................2

1.4. RESEARCH AIM AND OBJECTIVE..............................................................................2

1.5. SIGNIFICANCE OF THE RESEARCH/STUDY............................................................2

1.6. SCOPE OF THE RESEARCH/STUDY............................................................................3

1.7. LIMITATIONS OF THE RESEARCH/STUDY..............................................................4

1.8. RESEARCH STRUCTURE...............................................................................................4

1.9. OPERATIONAL DEFINITION OF TERMS...................................................................5

CHAPTER TWO...........................................................................................................................7

LITERATURE REVIEW.............................................................................................................7

2.1. INTRODUCTION.............................................................................................................7

VII
2.2. IMPORTANCE OF SENTIMENT ANALYSIS IN COURSE EVALUATION..........8

2.3. INNOVATION IN SENTIMENT ANALYSIS METHODS..........................................8

2.4. RELATED WORKS OF COURSE EVALUATION USING SENTIMENT


ANALYSIS.........................................................................................................................9

2.5. ASPECT BASED SENTIMENTS..................................................................................11

2.6. RELATED WORKS OF ASPECT BASED SENTIMENTS.......................................12

2.7. RESEARCH GAP IN PAST WORKS...........................................................................15

CHAPTER THREE.....................................................................................................................17

RESEARCH METHODOLOGY...............................................................................................17

3.1. INTRODUCTION TO DESIGN SCIENCE RESEARCH METHODOLOGY..........17

3.2. EVALUATION OF THE ARTIFACT............................................................................18

3.3. ASPECTS CONSIDERED IN RESEARCH...................................................................18

3.4. INTEGRATION OF AGILE SDLC................................................................................19

3.4.1. PHASES OF AGILE USED IN THE ARTIFACT DEVELOPMENT.................20

3.5. ANALYSIS OF EXISTING SYSTEM.............................................................................21

3.5.1. ANALYSIS OF EXISTING SYSTEMS: INCORPORATING DYNAMIC


SEMANTICS INTO PRE-TRAINED LANGUAGE MODEL FOR ASPECT-BASED
SENTIMENT ANALYSIS...................................................................................................22

3.5.2. ANALYSIS OF EXISTING SYSTEMS: BERT WITH AUXILIARY


SENTENCES FOR ABSA...................................................................................................24

3.5.3 LIMITATIONS OF EXISTING SYSTEMS.............................................................25

3.6. PROPOSED SYSTEM......................................................................................................27

3.7. REQUIREMENTS OF THE PROPOSED SYSTEM....................................................31

3.7.1. FUNCTIONAL REQUIREMENTS.........................................................................31

3.7.2. NON-FUNCTIONAL REQUIREMENTS...............................................................32

3.8. CONCLUSION..................................................................................................................32

VIII
CHAPTER FOUR.......................................................................................................................33

SYSTEM DESIGN AND IMPLEMENTATION......................................................................33

4.1 INTRODUCTION..............................................................................................................33

4.2 DEVELOPMENT TOOL, TECHNIQUES AND ENVIRONMENT............................33

4.3. DESIGN AND MODELLING OF THE SYSTEM.........................................................34

4.3.1. SYSTEM USE CASE DIAGRAM............................................................................35

4.3.2. SYSTEM ACTIVITY DIAGRAM............................................................................36

4.3.3. SYSTEM CLASS DIAGRAM...................................................................................38

4.4. AI PROCESS.....................................................................................................................38

4.5. BACKEND PROCESS......................................................................................................44

4.6. FRONTEND PROCESS...................................................................................................47

4.6.1. TOOLS AND TECHNOLOGIES USED.................................................................47

4.6.2. INTEGRATION WITH BACKEND........................................................................47

4.6.3. FUNCTIONALITIES...............................................................................................48

4.6.4. USER INTERFACE DESIGN..................................................................................48

4.6.5. DEPLOYMENT.........................................................................................................49

4.7. TESTING AND EVALUATION......................................................................................49

4.7.1 MODEL EVALUATION............................................................................................49

4.7.2 BACKEND TESTING: UNIT TESTING ON ASPECT EVALUATION


EXTRACTION FOR COURSE EVALUATION..............................................................56

4.7.3 FRONTEND TESTING..............................................................................................59

4.8 DISCUSSIONS OF FINDINGS.........................................................................................60

CHAPTER FIVE.........................................................................................................................62

SUMMARY AND CONCLUSION............................................................................................62

5.1 SUMMARY.........................................................................................................................62

IX
5.2 CONTRIBUTION OF RESEARCH AND CONCLUSION...........................................62

5.3 RECOMMENDATIONS...................................................................................................63

5.4 AREA FOR FURTHER RESEARCH..............................................................................63

X
LIST OF FIGURES

FIGURE 3.1: STEPS IN THE DESIGN SCIENCE RESEARCH METHODOLOGY........17


FIGURE 3.2: LIFECYCLE STAGES OF AGILE METHODOLOGY ILLUSTRATED IN
A DOUGHNUT CHART............................................................................................................20
FIGURE 4.1: INTERACTION BETWEEN THE FINE-TUNED MODEL, BACKEND,
FRONTEND, AND DATABASE COMPONENTS IN THE PROPOSED SYSTEM
ARCHITECTURE.......................................................................................................................28
FIGURE 4.2: USE CASE DIAGRAM FOR ASPECT BASED SENTIMENT ANALYSIS
COURSE EVALUATION SYSTEM.........................................................................................35
FIGURE 4.3: ACTIVITY FLOW DIAGRAM OF ASPECT-BASED SENTIMENT
ANALYSIS IN UNIVERSITY AND STUDENT PORTALS..................................................37
FIGURE 4.4: CLASS DIAGRAM OF ASPECT-BASED SENTIMENT ANALYSIS IN
UNIVERSITY AND STUDENT SYSTEM...............................................................................38
FIGURE 4.5: IMAGE SHOWING THE PRE-PROCESSED JSONL DATA FOR FINE-
TUNING.......................................................................................................................................40
FIGURE 4.6: COMPLETED FINE-TUNE STATUS UPDATE PROVIDED BY OPEN AI
CONTAINING TRAINED TOKENS, NUMBER OF EPOCHS AND TRAINING AND
VALIDATION LOSS..................................................................................................................42
FIGURE 4.7: STUDENT DATABASE SCHEMA...................................................................45
FIGURE 4.8: INFERENCE POINT INTERACTING WITH THE FINE-TUNED MODEL
.......................................................................................................................................................46
FIGURE 4.9: UNIVERSITY DASHBOARD PORTAL..........................................................48
FIGURE 4.10: EVALUATION METRICS SHOWING TRAINING LOSS AND
VALIDATION LOSS FOR FINE-TUNED MODEL AT DIFFERENT STEPS...................51
FIGURE 4.11: TERMINAL SHOWING THAT THE TEST WAS SUCCESSFUL............56
FIGURE 4.12: CHROME LIGHTHOUSE RESULT ON THE ABSA FOR COURSE
EVALUATION PLATFORM....................................................................................................59

XI
LIST OF TABLES

TABLE 2.1.:RELATED WORKS OF COURSE EVALUATION USING SENTIMENT


ANALYSIS 9
TABLE 2.2.:RELATED WORKS OF ASPECT BASED SENTIMENTS---------------------12
TABLE 4.1: ASPECT-BASED SENTIMENT ANALYSIS RESULTS COMPARISON- - -54
TABLE 4.2: EVALUATION METRICS FOR ASPECT-BASED SENTIMENT ANALYSIS
MODEL------------------------------------------------------------------------------------------------------55
TABLE 4.3: TEST CASE DESCRIPTION AND PROCEDURE FOR
GETCOURSEEVALUATIONASPECTS-------------------------------------------------------------57

XII
LIST OF APPENDIX

APPENDIX--------------------------------------------------------------------------------------------------69
APPENDIX 1------------------------------------------------------------------------------------------------69
ETHICAL CONSIDERATIONS IN USING ASPECT-BASED SENTIMENT ANALYSIS
FOR COURSE EVALUATION-------------------------------------------------------------------------69
APPENDIX 2------------------------------------------------------------------------------------------------71
USER INTERFACE DESIGN---------------------------------------------------------------------------71
APPENDIX 3 SOURCE CODE-------------------------------------------------------------------------74

XIII
CHAPTER ONE

INTRODUCTION
1.1. BACKGROUND OF THE STUDY

Course evaluation is an integral part of the educational process, allowing institutions to gather
valuable feedback from students about their learning experiences. Traditional methods often rely
on numerical ratings and open-ended comments. However, these comments are diverse and can
cover various aspects of the course. Analyzing them manually is time-consuming and
challenging due to their multifaceted nature. Aspect-based Sentiment Analysis (ABSA) offers a
solution by automatically identifying specific aspects and sentiments in student feedback
(Pontiki et al., 2014; Wang et al., 2020; Zhang et al., 2022), enabling educators to gain deeper
insights into the strengths and weaknesses of courses.

Student feedback holds the key to enhancing the quality of education. Understanding student’s
sentiments and opinions about different aspects of a course, such as teaching methods, course
content, assessments etc., can provide valuable insights to educators. Traditional course
evaluation methods often fail to capture the nuances of these multifaceted sentiments. ABSA, an
emerging field in Natural Language Processing (NLP), offers a more fine-grained approach to
analyzing student feedback by identifying aspects and associated sentiments within comments
(Zhang et al., 2022).

1.2. STATEMENT OF THE PROBLEM

The traditional manual analysis of student feedback comments in course evaluations poses
challenges due to the diverse nature of the comments. Educators often struggle to extract
valuable insights from the sheer volume of feedback. This study aims to address this problem by
leveraging Aspect-based Sentiment Analysis (ABSA) to automatically identify the aspects within
student feedback and associate sentiment polarities with them. The goal is to provide educators
with a more structured and comprehensive understanding of student opinions, enabling targeted
interventions for course improvement. By building upon the works of Pontiki et al. (2014), Wang
et al. (2020), Zhang et al. (2022), and other researchers in the field of aspect-based sentiment
analysis, this research aims to contribute to the application of ABSA in the context of course

XIV
evaluation. The proposed methodology will integrate ABSA techniques to analyze student
feedback, identify aspects, and provide sentiment insights that can aid educators in making
informed decisions to enhance course quality. Through the seamless integration of NLP and
education, this study intends to bridge the gap between raw student feedback and actionable
recommendations for educational institutions.

1.3. RESEARCH QUESTIONS

I. How can Aspect-based Sentiment Analysis (ABSA) be effectively applied to


enhance the analysis of course evaluations and reveal multifaceted student
feedback?
II. What are the specific aspects of student feedback in course evaluations, and how
do sentiment polarities associated with these aspects provide insights into the
strengths and weaknesses of courses?
III. How do educational institutions get the students to submit evaluation for a course,
how do students submit the evaluation and how does the aspects from the
evaluation get conveyed over to the educational institutions.

1.4. RESEARCH AIM AND OBJECTIVE

The aim of this study is to enhance the analysis of course evaluations by leveraging Aspect-
based Sentiment Analysis (ABSA) to reveal multifaceted student feedback, thereby providing
educators with deeper insights into the strengths and weaknesses of courses. The objectives are:

I. To review existing literature on course evaluations, and Aspect-based Sentiment Analysis


(ABSA) to establish a comprehensive understanding of the field.
II. To implement and fine-tune capable LLM to automatically identify specific aspects and
sentiment polarities in student feedback related to course evaluations.
III. To apply the ABSA techniques to the collected dataset of student feedback and analyze
the results in terms of aspect identification and sentiment distribution.

1.5. SIGNIFICANCE OF THE RESEARCH/STUDY

The significance of this study lies in its potential to revolutionize the course evaluation process.
By harnessing the power of natural language processing and sentiment analysis, educators can

XV
gain a nuanced understanding of student perceptions beyond quantitative ratings. The
significance includes:

I. Enhanced Insights: ABSA enables educators to pinpoint specific aspects of courses that
resonate with students and areas that need improvement, leading to more targeted
pedagogical enhancements.
II. Efficiency: The automation provided by ABSA streamlines the analysis of large volumes
of student feedback, saving time and effort compared to manual evaluation.
III. Holistic Evaluation: Traditional methods often miss intricate sentiments conveyed
through textual feedback. ABSA offers a comprehensive view of student opinions by
considering both aspects and sentiments.
IV. Informed Decision-making: Educators can make data-driven decisions to refine course
content, teaching methods, and overall curriculum based on multifaceted insights derived
from ABSA.
V. Academic Advancement: This study contributes to the field of educational technology by
demonstrating the applicability of ABSA in the context of course evaluations, opening
doors for further research and innovation.
VI. Educational Improvement: The findings and recommendations of this study can lead to
improvements in course design, instruction, and student engagement, ultimately
enhancing the quality of education.

1.6. SCOPE OF THE RESEARCH/STUDY

This study focuses on enhancing the analysis of course evaluations using Aspect-based
Sentiment Analysis (ABSA) techniques. The scope includes:

I. Collection and preprocessing of course evaluation data, including open-ended comments.


II. Implementation and customization of Large Language Models (LLMS) to identify
specific aspects and sentiment polarities in student feedback.
III. Creating a platform for integrating ABSA into the course evaluation process and utilizing
the multifaceted feedback.

XVI
1.7. LIMITATIONS OF THE RESEARCH/STUDY

Firstly, the data collected represents a single snapshot in time, reflecting the course evaluations
within a specific period. Consequently, the findings may not fully capture the dynamic nature of
student feedback over time. Moreover, the reliance on a single study introduces potential
limitations in establishing the long-term validity and generalizability of the results.

Furthermore, while this study aimed to comprehensively analyze student feedback using Aspect-
based Sentiment Analysis (ABSA), the scope of analysis may not encompass all dimensions of
student perceptions. Despite efforts to identify and incorporate relevant aspects and sentiments,
there may exist nuances and subtleties in student feedback that were not captured in this analysis.

Additionally, the subjectivity inherent in sentiment analysis introduces the possibility of biases
or inconsistencies in interpretation. While efforts were made to mitigate these biases through
rigorous annotation and validation processes, the subjective nature of sentiment analysis should
be acknowledged.

Lastly, the study's focus on the educational sector means that the findings may not be directly
applicable to other industries or domains. Extrapolation of the results beyond the scope of this
study should be approached with caution.

1.8. RESEARCH STRUCTURE

This project work is organized into five chapters. Chapter one introduces the concept of Aspect-
based Sentiment Analysis (ABSA) and presents a background of the work. It discusses the
problems and challenges faced by most existing methods of analyzing student feedback in course
evaluations, highlighting the need for more effective approaches such as ABSA.

Chapter two delves into the literature review of former work carried out on the topic. It provides
an in-depth analysis of previous research studies related to course evaluations, sentiment
analysis, and ABSA techniques. This chapter aims to establish a comprehensive understanding of
the field and identify gaps in existing literature that this research intends to address.

Chapter three furnishes a comprehensive analysis of the research methodology utilized in this
study, it intricately outlines the procedural steps involved in the research, spanning from the
design science methodology to the SDLC approach taken.

XVII
Chapter four delves into the implementation of Aspect-based Sentiment Analysis (ABSA)
techniques in analyzing student feedback.

Finally, chapter five contains the summary, conclusion, and recommendations drawn from the
findings of this research. The document is rounded off with the references and appendix
consisting of code used in the implementation of ABSA techniques, providing readers with
additional resources and technical details related to the research.

1.9. OPERATIONAL DEFINITION OF TERMS

In this study, several key terms are used extensively to describe concepts and methodologies. To
ensure clarity and consistency in their interpretation, the following operational definitions are
provided:

I. Course Evaluation: Refers to the systematic process of assessing and analyzing


various aspects of a course. This includes evaluating teaching effectiveness, course
content, assessment methods, and the overall learning experience. Course evaluations
often involve collecting feedback from students through surveys, questionnaires, or
other assessment instruments (Ameer & Hashim, 2020).
II. Aspect-based Sentiment Analysis (ABSA): A subfield of Natural Language
Processing (NLP) that focuses on identifying specific aspects or attributes within text
data and analyzing the sentiment or opinion associated with each aspect. ABSA
provides a fine-grained understanding of sentiment expressions by considering
contextually relevant aspects within the text (Xu et al., 2020; Sun, Huang, & Qiu,
2019).
III. Sentiment: Refers to the emotional orientation or tone expressed towards a specific
aspect or entity within textual data. Sentiment polarity is commonly categorized into
distinct classes, such as positive, negative, neutral, or nuanced emotions, which
collectively portray the overarching sentiment conveyed by the text (Zhang, Wang, &
Liu, 2018; Yadav & Vishwakarma, 2020).
IV. Fine-tuning: Refers to the process of adapting a pre-trained machine learning model
to a specific task or domain by further training it on task-specific data. Fine-tuning
involves adjusting the model's parameters to optimize its performance for the target

XVIII
task, such as ABSA for course evaluation analysis (Howard & Ruder, 2018; Devlin et
al., 2019).

XIX
CHAPTER TWO

LITERATURE REVIEW
2.1. INTRODUCTION
In recent years, online course learning has gradually become the mainstream of education,
reshaping traditional learning approaches and expanding educational accessibility (Pu et al.,
2021). As the demand for online courses grows, the quality of these courses becomes a critical
concern. Users' comments have emerged as a valuable source of feedback for improving online
courses, making sentiment analysis a pivotal tool for understanding user experiences (Pu et al.,
2021; Jiao et al., 2021). The sentiment expressed in user comments serves as a crucial indicator
of user satisfaction and provides insights into various aspects of course content, structure, and
delivery (Zhao et al., 2020). Positive sentiments often indicate effective teaching methods,
engaging content, and enjoyable learning experiences, while negative sentiments may highlight
areas requiring improvement (Pu et al., 2021). However, analyzing the sentiments of a large
volume of unstructured textual data poses challenges, necessitating innovative methods and
technologies (Lalata et al., 2019). To address these challenges, researchers have advanced
sentiment analysis methodologies. Pu et al. (2021) proposed an ensemble model that combines
Word2Vec and Glove for word vector representation, coupled with bidirectional long and short-
term networks and convolutional neural networks for feature extraction. Their multi-objective
gray wolf optimization (MOGWO) ensemble method demonstrated sentiment recognition
accuracy exceeding 91%, showcasing the potential of deep learning techniques in sentiment
analysis. Jiao et al. (2021) introduced a graph learning-based approach, leveraging graph
convolution networks (GCN) to capture syntactic dependencies between words and aspects of
sentences. This aspect-level sentiment classification method allowed for nuanced sentiment
analysis and achieved an accuracy of over 80% in sentiment classification, highlighting the
benefits of considering contextual relationships in sentiment analysis. Zhao et al. (2020)
harnessed machine learning, particularly the Support Vector Machine (SVM) model, to
transform evaluation text into quantifiable scores, enhancing the value of traditional teaching
evaluation systems. Their use of Flask to build a network project facilitated frontend data display
and interface design, facilitating user-friendly access to sentiment analysis results (Zhao et al.,

XX
2020). In this rapidly evolving landscape, the integration of sentiment analysis with education
evaluation systems has the potential to revolutionize course design and delivery, providing
educators with data-driven insights for continuous improvement (Lundqvist et al., 2020).

2.2. IMPORTANCE OF SENTIMENT ANALYSIS IN COURSE


EVALUATION
The sentiment embedded in user comments provides a unique perspective on the quality and
effectiveness of online courses. Understanding whether users' sentiments are positive, negative,
or neutral can be transformative for educators and institutions, guiding them in making informed
decisions to enhance the learning experience. This sentiment-based approach is particularly
crucial due to the subjective nature of educational experiences and the diverse factors that
contribute to user satisfaction.

Harnessing sentiment analysis can aid in understanding not only what users are saying but also
how they feel about various aspects of a course. Positive sentiments can highlight successful
teaching methods, engaging content, and valuable learning experiences. Conversely, negative
sentiments may reveal areas of improvement, such as confusing content, inadequate teaching
techniques, or technical issues. Furthermore, sentiment analysis can uncover nuanced insights
that traditional numerical ratings might miss, providing a comprehensive understanding of user
experiences.

2.3. INNOVATION IN SENTIMENT ANALYSIS METHODS


The advancement of sentiment analysis methodologies has paved the way for more accurate and
insightful evaluations of course feedback. Researchers have explored various techniques, such as
deep learning models, ensemble methods, and graph-based approaches. For instance, Pu et al.
(2021) introduced an ensemble deep learning model that leverages Word2Vec, Glove, and
advanced neural networks to extract deep features for sentiment analysis. The use of multi-
objective gray wolf optimization (MOGWO) enhanced model performance, yielding sentiment
recognition accuracy exceeding 91%.

Jiao et al. (2021) delved into aspect-level sentiment classification using a graph convolution
network (GCN). Their innovative approach captures the syntactic dependencies between words
and aspects of sentences, facilitating emotion classification for non-adjacent Chinese words.

XXI
Their system achieved an accuracy of over 80% in sentiment classification, demonstrating the
efficacy of their graph-based approach.

Zhao et al. (2020) harnessed machine learning, particularly the Support Vector Machine (SVM)
model, for sentiment analysis of course evaluation data. Their system transformed evaluation text
into quantifiable scores using SVM and applied a Flask-based interface for data display. This
integration of sentiment analysis and machine learning exemplifies the modernization of course
evaluation techniques.

2.4. RELATED WORKS OF COURSE EVALUATION USING


SENTIMENT ANALYSIS
Table 2.1.: Related works of course evaluation using sentiment analysis
Study Problem Method Used Solution Results
Identified
Pu et al. Quality Ensemble Achieved sentiment Higher accuracy
(2021) improvement model analysis using than other models,
based on combining ensemble model, stable results across
online course Word2Vec, F1score > 91% emotion levels
evaluations GloVe, CNN,
and MOGWO
Jiao et al. Sentiment Aspect-level Analyzed course Model accuracy >
(2021) analysis for sentiment evaluations with focus 80%, user-friendly
Chinese classification on course and teacher visual interface
course with GCN aspects designed
evaluations
Zhao et al. Sentiment SVM model Quantified sentiment Improved
(2020) analysis of with data in teaching evaluations understanding of
course preprocessing using SVM and DAG students' satisfaction
evaluation and Flask structure with course teachers
data frontend
Wu et al. Evaluating Dictionary- Used sentiment Provided evidence-
(2022) student based analysis on student based guidance for

XXII
learning sentiment reflections to guide course enhancement
through analysis for course improvement based on sentiment
sentiment learning analysis
analysis evaluation
Onan Sentiment Ensemble Developed sentiment Deep learning
(2020) analysis of learning and classification scheme outperformed other
MOOC deep learning for MOOC reviews methods, achieved
reviews for sentiment using deep learning accuracy of 95.80%
classification
Lundqvist Automated Sentiment Analyzed sentiment in Sentiment in posts
et al. (2020) sentiment analysis on MOOC posts, reflected feedback
analysis for MOOC posts categorized statements,
MOOC with different participants by prior beginners more
student prior knowledge positive
experience knowledge
Suprianto Information Generalized Combined GVSM and Achieved search and
et al. (2020) retrieval and Vector Space NBC for lecturer sentiment analysis
sentiment Model and evaluation through with proper
analysis for Naive Bayes sentiment analysis precision and
lecturers Classifier accuracy
Lalata et al. Sentiment Ensemble of Ensemble approach for Ensemble system
(2019) analysis for Naive Bayes, sentiment analysis of outperformed
faculty Logistic teacher evaluation individual classifiers
comment Regression, comments with 90.32%
evaluation SVM, accuracy
Decision Tree

2.5. ASPECT BASED SENTIMENTS


Sentiment analysis is increasingly viewed as a vital task both from an academic and a
commercial standpoint. The majority of current approaches, however, attempt to detect the
overall polarity of a sentence, paragraph, or text span, irrespective of the entities mentioned (e.g.,

XXIII
laptops) and their aspects (e.g., battery, screen) (Pontiki et al., 2014). SemEval2014 Task 4
aimed to foster research in the field of aspect-based sentiment analysis, where the goal is to
identify the aspects of given target entities and the sentiment expressed for each aspect. The task
provided datasets containing manually annotated reviews of restaurants and laptops, as well as a
common evaluation procedure. It attracted 163 submissions from 32 teams. Aspect-based
sentiment analysis aims to determine the sentiment polarity towards a specific aspect in online
reviews. Most recent efforts adopt attention-based neural network models to implicitly connect
aspects with opinion words. However, due to the complexity of language and the existence of
multiple aspects in a single sentence, these models often confuse the connections. As an
important fine-grained sentiment analysis problem, aspect-based sentiment analysis (ABSA),
aiming to analyze and understand people's opinions at the aspect level, has been attracting
considerable interest in the last decade. To handle ABSA in different scenarios, various tasks are
introduced for analyzing different sentiment elements and their relations, including the aspect
term, aspect category, opinion term, and sentiment polarity. Unlike early ABSA works focusing
on a single sentiment element, many compound ABSA tasks involving multiple elements have
been studied in recent years for capturing more complete aspect-level sentiment information.
However, a systematic review of various ABSA tasks and their corresponding solutions is still
lacking, which we aim to fill in this survey. More specifically, we provide a new taxonomy for
ABSA which organizes existing studies from the axes of concerned sentiment elements, with an
emphasis on recent advances of compound ABSA tasks.

ABSA aims to identify the aspect terms, their corresponding sentiment polarities, and the
opinion terms. ABSA predicts sentiment polarity towards a specific aspect in the given sentence.
While pre-trained language models such as BERT have achieved great success, incorporating
dynamic semantic changes into ABSA remains challenging. To this end, Zhang et al. (2022)
proposed to address this problem by Dynamic Re-weighting BERT (DR-BERT), a novel method
designed to learn dynamic aspect-oriented semantics for ABSA. Specifically, they first take the
Stack-BERT layers as a primary encoder to grasp the overall semantic of the sentence and then
fine-tune it by incorporating a lightweight Dynamic Re-weighting Adapter (DRA). Note that the
DRA can pay close attention to a small region of the sentences at each step and re-weigh the
vitally important words for better aspect-aware sentiment understanding. Finally, experimental

XXIV
results on three benchmark datasets demonstrate the effectiveness and the rationality of our
proposed model and provide good interpretable insights for future semantic modeling

ABSA involves three fundamental subtasks: aspect term extraction, opinion term extraction, and
aspect-level sentiment classification. Early works only focused on solving one of these subtasks
individually. Some recent work focused on solving a combination of two subtasks, e.g.,
extracting aspect terms along with sentiment polarities or extracting the aspect and opinion terms
pair-wisely. More recently, the triple extraction task has been proposed, i.e., extracting the
(aspect term, opinion term, sentiment polarity) triples from a sentence. However, previous
approaches fail to solve all subtasks in a unified end-to-end framework. Mao et al. (2021)
proposed a complete solution for ABSA. They constructed two machines reading comprehension
(MRC) problems, and solve all subtasks by joint training two BERT-MRC models with
parameters sharing. They conducted experiments on these subtasks and results on several
benchmark datasets demonstrate the effectiveness of their proposed framework, which
significantly outperforms existing state-of-the-art methods

2.6. RELATED WORKS OF ASPECT BASED SENTIMENTS


Table 2.2.: Related works of Aspect based sentiments
Paper Problem Method Used Proposed Results
Identified Solution
Pontiki et Lack of aspect- Manual SemEval- Attracted 163
al. (2014) based sentiment annotation of 2014 Task 4 submissions from 32
analysis reviews, with teams
approaches common annotated
evaluation restaurant
procedure and laptop
reviews
Wang et Confusing Relational Graph R-GAT using Improved connections,
al. (2020) aspect-opinion Attention syntax significant GAT
connections in Network (R- information performance
neural models GAT) and new tree improvement
structure
Zhang et Lack of Survey and Taxonomy Summarized ABSA

XXV
al. (2022) systematic classification of for ABSA tasks, discussed
review for ABSA tasks and tasks, solutions and challenges
ABSA tasks solutions emphasis on
and solutions compound
tasks
Pontiki et Continuation of Provided SemEval- Attracted 245
al. (2016) ABSA tasks for training and 2016 Task 5 submissions from 29
multiple testing datasets with 19 teams
languages and in multiple training and
domains languages and 20 testing
domains datasets
Xu et al. Need for Post-training on Novel post- Improved performance
(2019) review-based BERT for training in RRC and ABSA
reading Review Reading approach on
comprehension Comprehension BERT for
for ABSA (RRC) RRC
Yan et al. Lack of unified Unified BART model Substantial performance
(2021) framework for generative to solve all gain, unified solution
ABSA subtasks formulation with ABSA for subtasks
BART subtasks
Zhang et Need for Dynamic Re- Novel Effective model with
al. (2022) dynamic aspect- weighting BERT method to better sentiment
oriented (DR-BERT) learn understanding
semantics in dynamic
ABSA aspect-
oriented
semantics
Sun et al. Enhancing Constructing BERT-based State-of-the-art results
(2019) ABSA using auxiliary approach on SentiHood and
auxiliary sentence and with SemEval-2014
sentence and fine-tuning auxiliary
BERT BERT sentence
construction
Peng et Lack of Aspect Two-stage Outperformed baselines,

XXVI
al. (2019) complete sentiment triplet framework complete solution for
solution for extraction for ASTE ABSA
ABSA (ASTE)
Mao et Need for joint Joint Training Two BERT- Superior performance
al. (2021) training Dual-MRC MRC models compared to existing
framework for Framework with for joint methods
ABSA BERT training
He et al. Lack of joint Interactive Joint learning Outperformed baselines
(2019) information Multi-Task of tasks at on benchmark datasets
integration in Learning token and
ABSA Network (IMN) document
levels
He et al. Addressing Meta-based Self- Meta- Competitive results on
(2022) biases and training with weighter to ABSA with MSM
imbalances in meta-weighter balance
ABSA (MSM) learning and
avoid biases
Fei et al. Integrating Hierarchical Multiplex Improved performances
(2022) ABSA subtasks dependency and cascade on subtasks with
using multiplex multiplex framework multiplex strategy
cascade decoding for unified
ABSA
D'aniello Clarifying KnowMIS- Different Improved analysis using
et al. differences in ABSA model for tools and separate metrics and
(2022) ABSA separate metrics for techniques
dimensions dimensions sentiment,
affect,
emotion,
opinion

2.7. RESEARCH GAP IN PAST WORKS


When comparing the related works in the fields of course evaluation and aspect-based sentiment
analysis, a clear research gap emerges in the integration of these two areas, particularly in the
context of enhancing course evaluation using aspect-based sentiment analysis for revealing

XXVII
multifaceted student feedback. While the related works showcase advancements in sentiment
analysis techniques and their applications, the following research gap is evident

1. Integration of Aspect-based Sentiment Analysis in Course Evaluation:

While the provided related works cover various aspects of sentiment analysis, none seem to
directly address the integration of aspect-based sentiment analysis into the course evaluation
process to reveal nuanced student feedback across different dimensions. Most studies in the
related works focus on sentiment analysis in broader contexts, such as product reviews, MOOC
posts, or faculty evaluations. However, the proposed approach in the abstract aims to dissect
student feedback into distinct facets like teaching style, content relevance, and assessment
methods. This gap highlights the need for research that specifically explores the benefits and
challenges of incorporating aspect-based sentiment analysis techniques into the course evaluation
process.

2. Multidimensional Insights into Student Feedback:

Existing related works largely emphasize sentiment classification or sentiment polarity detection.
However, the abstract highlights the goal of categorizing student feedback into specific
dimensions that go beyond a binary positive/negative sentiment. This implies a need for research
that develops methods capable of identifying and analyzing various aspects or dimensions within
student feedback and uncovering nuances within those aspects. These multidimensional insights
would provide educators with a more comprehensive understanding of both strengths and areas
for improvement.

3. Effectiveness of Aspect-based Sentiment Analysis for Curriculum Enhancement:

While sentiment analysis techniques have been successfully applied in various domains, the
related works do not explicitly address the potential impact of integrating aspect-based sentiment
analysis into course evaluation on curriculum enhancement and instructional refinement. The
research suggests that this approach could enable data-driven decision-making for such
enhancements. A research gap thus exists in empirically investigating whether the insights
derived from aspect-based sentiment analysis indeed lead to meaningful improvements in
curriculum design and teaching strategies, ultimately influencing student outcomes.

XXVIII
XXIX
CHAPTER THREE

RESEARCH METHODOLOGY
This chapter presents the research methodology utilized in crafting a platform intended to
transform the landscape of course evaluations within educational institutions through Aspect-
based Sentiment Analysis (ABSA). Through a cohesive integration of Design Science
Methodology (DSM) with Agile Software Development Life Cycle (SDLC) principles, the aim
is to effectively tackle significant challenges and attain research objectives with accuracy and
efficiency.

3.1. INTRODUCTION TO DESIGN SCIENCE RESEARCH


METHODOLOGY

Figure 3.1: Steps in the Design Science Research Methodology

Design Science Research Methodology (DSRM) serves as the cornerstone of this research
endeavor. Rooted in the creation and evaluation of innovative artifacts, DSRM offers a
structured framework for solving real-world problems. The platform, envisioned to facilitate

XXX
course evaluations in educational institutions, embodies the essence of DSRM by addressing
critical issues and contributing to the advancement of educational practices.

3.2. EVALUATION OF THE ARTIFACT


The platform underwent rigorous evaluation to ascertain its effectiveness and utility in enhancing
course evaluations. Utilizing evaluation metrics such as F1 score, the performance of the AI
model in extracting aspects from course reviews was assessed. These metrics ensured the
accuracy and reliability of the aspect extraction process. Additionally, the quality of the training
data used for the AI model was verified, ensuring it accurately represented real course reviews
from diverse sources.

While direct user feedback was not collected, insights gained from the evaluation process,
including the analysis of performance metrics and examination of training data, informed
iterative refinements aimed at enhancing the platform's functionality. Continual optimization and
refinement ensured the platform effectively meets the needs of educational institutions and
students alike, fostering a seamless and efficient course evaluation process.

3.3. ASPECTS CONSIDERED IN RESEARCH


A comprehensive analysis of various aspects critical for evaluating course effectiveness in
educational settings was conducted. The following aspects were meticulously examined:

I. Instructional Effectiveness: This aspect evaluates instructors' ability to convey course


material effectively, engage students, and facilitate meaningful learning experiences.

II. Learning Environment: Focused on the overall atmosphere and conditions within the
classroom or educational setting, this aspect considers factors such as classroom layout,
resources, and student interactions, which significantly influence learning outcomes.

III. Assessments and Feedback: This aspect examines the methods used to assess student
learning and provide feedback, including the fairness, validity, and timeliness of assessments, as
well as the effectiveness of feedback mechanisms in promoting student growth.

XXXI
IV. Assessment Evaluation Methods: Delving into the processes used to evaluate the
effectiveness of assessments, this aspect includes strategies for analyzing assessment data,
identifying areas for improvement, and adjusting assessment practices accordingly.

V. Lecture Quality: This aspect evaluates the quality of lectures delivered by instructors,
considering factors such as clarity, organization, engagement, and relevance to course objectives.

VI. Instructor Availability: Assessing the accessibility and responsiveness of instructors


outside of class, this aspect includes availability for office hours, responsiveness to emails or
messages, and willingness to provide support and guidance to students.

VII. Classroom Interaction: Exploring the level of interaction and engagement among students
and between students and instructors during class sessions, this aspect encompasses opportunities
for discussion, collaboration, and active participation.

Justification for Aspects Considered:

These aspects align closely with the comprehensive understanding of teaching effectiveness
highlighted in relevant literature. Studies such as Ajmal, Durrani, and Mohammad (2021)
emphasize the multidimensional nature of teaching effectiveness, encompassing instructional
competence, classroom management, and student engagement. By examining these facets of
course delivery and student experiences, the research aims to provide valuable insights for
enhancing teaching and learning outcomes in educational settings.

3.4. INTEGRATION OF AGILE SDLC


The Agile Software Development Life Cycle (SDLC) played a crucial role in the iterative
development and deployment of the platform. Each phase of SDLC, from requirements gathering
to deployment and maintenance, was meticulously incorporated into the research methodology.

XXXII
Figure 3.2: Lifecycle Stages of Agile Methodology Illustrated in a Doughnut Chart.

According to Al-Saqqa, Sawalha, and AbdelNabi (2020), Agile is a wide umbrella of software
development beliefs that involves a conceptual framework for software engineering, beginning
with a planning phase and progressing iteratively and incrementally towards deployment, aiming
to reduce overhead and facilitate adaptability to changes without risking the process or
necessitating excessive rework.

Agility, adaptability, and responsiveness to evolving user needs were maintained throughout,
ensuring seamless integration of AI-driven capabilities and enhanced course evaluations through
Aspect-based Sentiment Analysis.

3.4.1. PHASES OF AGILE USED IN THE ARTIFACT DEVELOPMENT


I. Requirements Gathering:

While traditional requirements gathering processes were not followed explicitly, the recognition
of the critical role of course evaluations in educational institutions prompted the need to connect
effectively with students. This understanding served as the foundation for determining the
platform's functionalities and features.

II. Planning:

In the planning phase, a comprehensive roadmap was outlined, emphasizing the iterative
development approach inherent in agile methodology. The focus was on prioritizing features that

XXXIII
would simplify course evaluations for both educational institutions and students, aligning with
the agile principle of incremental delivery.

III. Development:

The development process commenced with the refinement of the Aspect-based Sentiment
Analysis (ABSA) model based on real course review data. This iterative refinement ensured the
accuracy and effectiveness of the model in identifying aspects and sentiments within student
feedback. Subsequently, the backend and database were constructed, integrating seamlessly with
the fine-tuned ABSA model. The user interface was designed with simplicity and intuitiveness in
mind, reflecting the agile principle of responding to change over following a rigid plan.

IV. Testing:

Comprehensive testing was conducted throughout the development process such as using
postman to validate the platform's functionality via API endpoints and reliability and the use of
jasmine for testing on the server side. This iterative testing approach allowed for the early
identification and resolution of any issues, ensuring the quality of the final product.

V. Deployment:

Deployment of the platform was executed on cloud infrastructure, aligning with agile principles
of flexibility and scalability. This cloud-based deployment facilitated accessibility for
educational institutions and students, enabling seamless interaction with the platform from
anywhere and at any time.

VI. Maintenance:

As the platform entered its Minimum Viable Product (MVP) stage, ongoing maintenance became
essential. Feedback from users will inform iterative improvements to the platform, ensuring
continuous enhancement and optimization in line with agile principles.

3.5. ANALYSIS OF EXISTING SYSTEM


This section thoroughly examines existing methods of Aspect-Based Sentiment Analysis
(ABSA), focusing on how they handle the intricate sentiment variations related to different

XXXIV
aspects within text. By analyzing these approaches, we aim to gain valuable insights to enhance
and develop our own ABSA framework.

3.5.1. ANALYSIS OF EXISTING SYSTEMS: INCORPORATING DYNAMIC


SEMANTICS INTO PRE-TRAINED LANGUAGE MODEL FOR ASPECT-BASED
SENTIMENT ANALYSIS
We will explore a recent paper, "Incorporating Dynamic Semantics into Pre-Trained Language
Model for Aspect-based Sentiment Analysis Zhang et al. (2022)" and discuss its strengths and
weaknesses in relation to our proposed system.

This paper introduces DR-BERT, a novel method for ABSA that leverages pre-trained language
models (like BERT) and incorporates the concept of dynamic human semantic comprehension.

I. Research Question:

Can pre-trained language models be enhanced by incorporating dynamic semantics for improved
ABSA performance?

II. Methodology:

1. Input: The model receives a sentence-aspect pair (S, A) as input such as

Sentence (S): "The movie had a captivating plot, but the special effects were
disappointing."

Aspect (A): "special effects"

2. Embedding Module: Words from both the sentence and aspect are converted into numerical
representations using BERT embedding.

3. BERT Encoder: A pre-trained BERT model captures overall sentence meaning.

4. Dynamic Re-weighting Adapter (DRA): This is DR-BERT's core innovation. It iteratively


refines the sentence representation by:

Identifying the most relevant word in the sentence related to the aspect.

Focusing on this word and updating the overall sentence representation using a Gated
Recurrent Unit (GRU).

XXXV
5. Sentiment Prediction Module: Classifies the sentiment polarity (positive, negative, neutral) for
the identified aspect within the sentence.

III. Findings:

DR-BERT outperforms existing ABSA methods, demonstrating its effectiveness in capturing


dynamic semantics related to aspects. Additionally, the model offers interpretability by
highlighting the most relevant words for sentiment prediction.

IV. Strengths:

i. Addresses Dynamic Semantics: DR-BERT tackles a limitation of existing ABSA methods by


considering how humans dynamically focus on different parts of a sentence while
comprehending it in relation to a specific aspect.

ii. Improves BERT Performance: The DRA refines the sentence representation, potentially
leading to better ABSA performance compared to vanilla BERT.

iii. Interpretability: The model reveals the most crucial words for sentiment prediction in each
step.

V. Weaknesses:

i. Computational Cost: The iterative re-weighting process might require more computational
resources compared to simpler ABSA models.

ii. Limited Data Dependence: DR-BERT's effectiveness might be dependent on a large amount
of training data for the ABSA task.

iii. Hyper parameter Tuning: The number of re-weighting iterations needs careful tuning for
optimal performance.

VI. Conclusion

DR-BERT presents a valuable approach to ABSA by considering dynamic semantics. However,


our proposed which would be covered later in the chapter might be more versatile and potentially
achieve superior performance with sufficient data. The choice between these methods depends
on the specific needs and priorities of the ABSA task.

XXXVI
3.5.2. ANALYSIS OF EXISTING SYSTEMS: BERT WITH AUXILIARY SENTENCES
FOR ABSA

This section analyses an existing method for Aspect-Based Sentiment Analysis (ABSA) that
leverages pre-trained language models (LSTMs specifically) – Bidirectional Encoder
Representations from Transformers (BERT) (Liu et al., 2021). This approach addresses a key
challenge in ABSA: how to utilize BERT's strengths in sentence-pair classification for tasks
traditionally involving sentiment analysis of single sentences.

I. Methodology:
The core concept involves constructing "auxiliary sentences" related to the target aspect and
sentiment within a given sentence. This essentially transforms the single-sentence ABSA task
into a sentence-pair classification problem, which aligns well with BERT's capabilities. Here's a
breakdown of the BERT with auxiliary sentences approach:
1. Sentence Pair Classification with Auxiliary Sentences:
i. Auxiliary sentences are constructed to relate to the target aspect and sentiment within a
sentence (e.g., "What do you think about the special effects?" for aspect "special effects"
and sentiment "negative").
ii. This transforms the single-sentence ABSA task into a sentence-pair classification
problem, where the model needs to classify the relationship between the original sentence
and the auxiliary sentence regarding the target aspect and sentiment.

2. Fine-tuning BERT:
i. The pre-trained BERT model is fine-tuned on labeled ABSA data for this sentence-
pair classification task.
II. Findings:
Research by Liu et al. (2021) demonstrates state-of-the-art performance on benchmark datasets
using BERT-pair models with auxiliary sentences for both aspect detection and sentiment
classification. This success is attributed to BERT's strengths in sentence-pair classification tasks
due to its pre-training on masked language modeling and next sentence prediction.
Strengths:
i. Achieves state-of-the-art performance on benchmark datasets.

XXXVII
ii. Leverages BERT's strengths in sentence-pair classification for ABSA tasks.
Weaknesses:
i. Performance can be impacted by the design of auxiliary sentences (explored methods
include QA-M, NLI-M, QA-B, NLI-B).
ii. Requires computationally expensive fine-tuning of the BERT model.
iii. Relies on pre-trained models, which can be a limitation if fine-tuning data is scarce or
domain-specific (Tang et al., 2020).
III. Conclusion:

The BERT with auxiliary sentences approach offers a successful method for ABSA. However,
limitations exist in auxiliary sentence design and computational cost.

3.5.3 LIMITATIONS OF EXISTING SYSTEMS


This section analyzes the limitations of the two explored ABSA systems: DR-BERT and BERT
with auxiliary sentences. While both systems offer advancements in ABSA, they possess certain
limitations that can be addressed by our proposed approach utilizing fine-tuned GPT-3.5.

I. Limitations of DR-BERT

DR-BERT tackles the challenge of incorporating dynamic semantics into sentiment analysis.
However, several limitations can hinder its wider implementation:

i. Computational Cost: The core innovation of DR-BERT lies in its Dynamic Re-weighting
Adapter (DRA). This iterative process of refining the sentence representation can be
computationally expensive compared to simpler ABSA models. This increased computational
demand might become a hurdle for real-world applications with resource constraints.

ii. Data Dependence: The effectiveness of DR-BERT might be highly dependent on the
availability of a large amount of training data specific to the ABSA task. Limited training data
can hinder the model's ability to learn the intricate relationships between words and aspects for
accurate sentiment classification.

iii. Hyper parameter Tuning: DR-BERT relies on a critical hyperparameters: the number of re-
weighting iterations within the DRA. Fine-tuning this parameter is crucial for optimal

XXXVIII
performance. Finding the ideal number of iterations can be a complex task, requiring significant
experimentation and potentially limiting the model's generalizability.

II. Limitations of BERT with Auxiliary Sentences

The BERT with auxiliary sentences approach leverages the strengths of BERT for ABSA tasks.
However, it also possesses limitations that can be overcome by alternative approaches:

1. Auxiliary Sentence Design: The performance of this system can be significantly impacted by
the design of the auxiliary sentence. Different methods like QA-Matching (QA-M) or Natural
Language Inference Matching (NLI-M) are explored, and each might exhibit varying levels of
effectiveness depending on the specific task and data. Designing optimal auxiliary sentences can
be a complex process.

2. Computational Cost: Fine-tuning BERT, a pre-trained language model, can be


computationally expensive. This can limit the scalability and accessibility of the approach,
especially for resource-constrained environments.

3. Reliance on Pre-trained Models: This system relies on the effectiveness of pre-trained models
like BERT. If the fine-tuning data is scarce or domain-specific, the overall performance of the
system can be limited. BERT might not be specifically optimized for the nuances of sentiment
analysis within a particular domain, such as course evaluation analysis.

III. Limitations of Existing Techniques for Course Evaluation Analysis

While these existing techniques offer a foundation, they are hindered by limitations that can be
particularly impactful when applied to course evaluation analysis:

1. Keyword-based methods: These methods can be overly simplistic and miss contextual nuances
within student responses (Baker et al., 2010). For example, the keyword "difficult" might not
always convey negative sentiment (e.g., "a difficult but rewarding course").

2. Sentiment lexicon analysis: The effectiveness of this approach relies heavily on the accuracy
and completeness of the sentiment lexicon used (Mohammad & Turney, 2010). Lexicons might
not capture domain-specific language used in course evaluations, potentially leading to
misinterpretations (Hu & Liu, 2004).

XXXIX
3. Thematic analysis: While offering rich insights, thematic analysis is labor-intensive and time-
consuming, limiting its scalability for analyzing large datasets of open-ended responses (Nowell
et al., 2017).

These limitations highlight the need for advanced techniques that can effectively analyze open-
ended course evaluation responses while considering the specific context and domain of course
evaluation.

3.6. PROPOSED SYSTEM


This sub-section introduces a novel system utilizing fine-tuned GPT-3.5 for Aspect-Based
Sentiment Analysis (ABSA) of open-ended course evaluation responses. By addressing the
limitations identified in existing systems (Section 4.2.4), this approach aims to provide a more
comprehensive and efficient method for analyzing student feedback.

I. Rationale for Fine-tuned GPT-3.5

GPT-3.5, a large language model (LLM) pre-trained on a massive dataset of text and code, offers
several advantages for course evaluation analysis:

1. Contextual Understanding: GPT-3.5 excels at understanding the nuances of language,


including context and sentiment. This allows it to move beyond simple keyword matching and
capture the true meaning behind student responses in course evaluations (Yu et al., 2020).

2. Domain-Specific Fine-tuning: Fine-tuning GPT-3.5 on course evaluation data allows it to


adapt to the specific language and sentiment expressions used by students in this domain
(Howard & Ruder, 2018). This domain adaptation is crucial for accurate ABSA within the
context of course evaluation.

II. System Architecture

The proposed system architecture for the Aspect-Based Sentiment Analysis (ABSA) for course
evaluation web application comprises distinct components tailored to accommodate the
functionalities and requirements of both educational institutions (administrators) and students.
These components, along with the chosen technologies, facilitate seamless data flow and
interaction within the system

XL
Figure 4.1: Interaction between the fine-tuned model, backend, frontend, and database
components in the proposed system architecture

Figure 4.1 provides a visual representation of the proposed system architecture, delineating the
flow of data and processes across the backend, frontend, and database components, and their
interaction with the fine-tuned GPT-3.5 model.

I. Fine-tuned GPT-3.5 Model:

The fine-tuned GPT-3.5 model serves as the backbone of the Aspect Based Sentiment Analysis
(ABSA), adept at analyzing course evaluation feedback with precision and comprehensiveness.
Trained on a dataset curated from course evaluations, the model encapsulates nuanced
understandings of language expressions, enabling accurate sentiment classification across
various aspects of course feedback.

II. Backend Infrastructure (Node.js):

Developed using Node.js, the backend infrastructure forms the core computational engine of the
web application. It handles data processing, model inference, authentication, and communication
between the frontend and the database.

Node.js, with its non-blocking, event-driven architecture, ensures efficient handling of


concurrent requests, optimizing the system's performance and scalability.

III. Frontend Interface (React.js):

XLI
The frontend interface, built using React.js, caters to the diverse needs of both university
administrators and students. It features separate dashboards for administrators (university) and
students, each tailored to their respective roles and responsibilities.

University administrators access functionalities for data management, result visualization, and
administrative tasks, while students utilize the interface for submitting course evaluation forms.

IV. Database Management System (MongoDB):

MongoDB serves as the persistent data store, housing course evaluation data, user profiles, and
system configurations. Its flexible document-oriented architecture facilitates seamless storage
and retrieval of structured and unstructured data.

Leveraging Mongo DB’s scalability and performance capabilities, the system ensures robust data
management and high availability, accommodating the growing volume of course evaluations
and user interactions.

These components, meticulously integrated and orchestrated, enable the ABSA web application
to deliver a user-centric experience while maintaining efficiency, reliability, and scalability. The
collaboration between Node.js backend, MongoDB database, and React.js frontend empowers
educators and students alike to engage in informed decision-making and continuous
improvement initiatives within the academic realm.

III. Advantages of the Proposed System

The proposed system offers several advantages over existing techniques for course evaluation
analysis:

1. Improved Accuracy:

By leveraging GPT-3.5's capabilities in understanding context and sentiment, the system has the
potential to provide more accurate ABSA results compared to simpler techniques like keyword-
based analysis or sentiment lexicon analysis.

2. Domain Specificity:

Fine-tuning on course evaluation data allows the system to adapt to the specific language used
by students, leading to more accurate interpretations and sentiment classifications.

XLII
3. Scalability:

The use of a pre-trained LLM like GPT-3.5 offers potential for scalability, enabling efficient
analysis of large datasets of open-ended responses.

4. Reduced Computational Cost:

Utilizing cloud-based platforms for fine-tuning GPT-3.5 has the potential to reduce
computational cost compared to some traditional fine-tuning methods.

5. Potential for Enhanced Accuracy:

Through the integration of sophisticated language models and domain-specific fine-tuning, the
proposed system has the potential to offer more accurate and nuanced insights into course
evaluations compared to conventional manual assessment methods.

6. Efficiency and Scalability:

Automation enabled by the fine-tuned GPT-3.5 model streamlines the evaluation process,
significantly reducing the time and effort required for analysis. Moreover, the system's scalability
allows for the efficient handling of large volumes of course feedback data.

7. Comprehensive Analysis:

Unlike traditional methods that often focus on predefined evaluation criteria, the proposed
system can comprehensively analyze course feedback, capturing both explicit and implicit
sentiments across diverse aspects of the learning experience.

8. Promotion of Objectivity:

By minimizing human bias and subjectivity inherent in manual evaluation processes, the
automated nature of the proposed system promotes objectivity and consistency in assessing
course feedback, leading to more reliable outcomes.

XLIII
3.7. REQUIREMENTS OF THE PROPOSED SYSTEM
The requirements of the proposed Aspect-Based Sentiment Analysis (ABSA) system outline the
functionalities, constraints, and expectations that the system must meet to fulfill its intended
purpose of analyzing course evaluation feedback.

3.7.1. FUNCTIONAL REQUIREMENTS


I. Course Review Input Interface: Design an intuitive interface allowing students to submit
detailed course reviews, including text-based feedback and ratings.

II. University and Department Selection: Implement a user-friendly interface for students to
select their university and department, facilitating accurate course selection for review.

III. Forms Management: Develop an efficient system for categorizing and managing submitted
and un-submitted forms, enabling administrators to track the completion status of course
evaluations.

IV. Administrative Control: Provide administrators with comprehensive control over review
forms, including the ability to create, modify, adjust visibility settings, and remove forms as
necessary.

V. Feedback Visualization: Create visual representations, such as pie charts, to present


aggregated feedback data to administrators, allowing them to quickly grasp overall sentiment
trends and drill down into individual aspect-based reviews.

VI. Aspect-based Sentiment Analysis: Implement advanced sentiment analysis techniques,


focusing on specific aspects crucial to course delivery quality, such as instructional effectiveness,
learning environment, and assessment methods.

VII. Insight Generation: Develop algorithms to extract meaningful insights from sentiment
analysis results, categorizing feedback into positive, neutral, and negative sentiments for each
aspect of course delivery.

XLIV
3.7.2. NON-FUNCTIONAL REQUIREMENTS
I. Scalability: Ensure the system architecture can accommodate a growing volume of course
reviews and user interactions, with provisions for scaling resources dynamically to meet
increasing demands.

II. Security Measures: Implement robust data security protocols to safeguard user privacy and
maintain the confidentiality and integrity of sensitive information throughout the research
process.

III. Accuracy and Reliability: Validate sentiment analysis algorithms rigorously against ground
truth data to ensure high accuracy and reliability in identifying and categorizing sentiment
expressions in course reviews.

IV. Usability and User Experience: Conduct usability testing to refine the user interface and
optimize user experience, ensuring that both administrators and students can interact with the
system efficiently and effectively.

V. Performance Optimization: Employ efficient data processing techniques and optimize


system performance to minimize latency and deliver timely sentiment insights to users.

VI. Continuous Availability: Implement redundancy and failover mechanisms to ensure


uninterrupted access to the system, minimizing downtime and maximizing availability for users
engaged in research activities.

3.8. CONCLUSION
In conclusion, Chapter Three delineates a systematic approach to research methodology,
harmonizing Design Science Methodology with Agile Software Development Life Cycle
principles. The platform represents a tangible artifact crafted to address the complexities of
course evaluations in educational institutions, fostering enriched learning experiences and
institutional effectiveness. Through the meticulous application of DSRM and the integration of
SDLC, positive transformations in educational assessment practices are anticipated, ushering in a
new era of data-driven decision-making and student-focused education.

XLV
CHAPTER FOUR

SYSTEM DESIGN AND IMPLEMENTATION


4.1 INTRODUCTION
This chapter focuses on the design and implementation of the Aspect-Based Sentiment Analysis
(ABSA) system for course evaluation feedback. It outlines the process of translating theoretical
concepts into practical solutions, considering both academic rigor and real-world applicability.
The chapter covers various aspects including system architecture, backend processing, frontend
interface development, and integration of artificial intelligence techniques. Additionally, it
discusses system requirements, testing methodologies, and evaluation criteria to ensure the
effectiveness and reliability of the proposed solution.

4.2 DEVELOPMENT TOOL, TECHNIQUES AND ENVIRONMENT


This section outlines the development tools, techniques, and environment utilized in the project,
emphasizing the integration of various technologies to create a cohesive and efficient system for
analyzing course evaluation feedback using Aspect-Based Sentiment Analysis (ABSA).

I. React:

React, a popular JavaScript library for building user interfaces, was employed to develop
the front-end of the application. Its component-based architecture facilitated the creation
of a dynamic and responsive user interface, allowing for efficient rendering and updates
of course evaluation data.

II. Node.js:

Node.js served as the runtime environment for executing JavaScript on the server side. Its
non-blocking, event-driven architecture enabled the development of a highly scalable and
performant back-end, crucial for handling the asynchronous operations required in
sentiment analysis processing.

III. Express.js:

XLVI
Express.js, a minimal and flexible Node.js web application framework, was used to build
the RESTful API that connects the front-end with the back-end. It provided a robust set
of features for handling HTTP requests, middleware integration, and routing, ensuring
smooth communication between the client and server.

IV. MongoDB:

MongoDB, a NoSQL database, was chosen for its flexibility in handling unstructured
data. It stored the course evaluation comments and the corresponding sentiment analysis
results, allowing for efficient querying and retrieval of evaluation aspects.

V. Visual Studio Code:

Visual Studio Code, a versatile and powerful code editor, was the primary development
environment. Its extensive range of extensions and features, such as IntelliSense and
debugging tools, enhanced productivity and code quality.

VI. OpenAI:

OpenAI's language models were fine-tuned to perform the sentiment analysis and aspect
extraction tasks. The integration of these models enabled the identification of
multifaceted insights from the course evaluation comments, contributing to a more
nuanced understanding of student feedback.

4.3. DESIGN AND MODELLING OF THE SYSTEM


The design and modeling phase involves structuring the system's architecture and functionality
to meet the requirements of course evaluation analysis. This phase encompasses various
activities, including the development of use case diagrams, class diagrams, and other modeling
techniques to depict the system's behavior and interactions.

XLVII
4.3.1. SYSTEM USE CASE DIAGRAM
The use case diagram illustrates the various interactions between users and the aspect-based
course evaluation system. It outlines the system's functionalities and the actors involved in
utilizing those functionalities.

Figure 4.2: Use Case Diagram for aspect based sentiment analysis course evaluation system

The provided use case diagram (Figure 4.7) offers a graphical representation of the functional
interactions within the aspect-based sentiment analysis (ABSA) course evaluation system. This
analysis explores the actors involved and the functionalities they can leverage within the system.

XLVIII
4.3.2. SYSTEM ACTIVITY DIAGRAM

XLIX
Figure 4.3: Activity Flow Diagram of Aspect-Based Sentiment Analysis in University and
Student Portals

The activity diagram illustrates the dynamic process flow within the aspect-based sentiment
analysis (ABSA) course evaluation system, delineating the interactions and operations within
both the university and student portals.

The diagram captures the sequential activities involved in the evaluation process, starting from
the submission of course feedback by students to the comprehensive analysis and visualization of
results by university administrators. It elucidates the distinct pathways followed by students and
administrators within their respective portals, showcasing the nuanced functionalities tailored to
their roles and objectives.

Through a series of interconnected nodes and transitions, the activity diagram elucidates the
intricate orchestration of tasks encompassing data submission, preprocessing, sentiment analysis,
and result visualization. It provides a visual representation of the systematic workflow inherent
in the ABSA system, elucidating the underlying processes driving efficient course evaluation and
feedback mechanisms.

The delineation of activities within the university and student portals offers a holistic view of the
system's functionality, emphasizing its adaptability to diverse user roles and facilitating a

L
seamless exchange of information between stakeholders. Overall, the activity diagram serves as a
comprehensive roadmap for understanding the operational dynamics of the ABSA course
evaluation system, underscoring its efficacy in enhancing the quality and responsiveness of
educational feedback mechanisms.

4.3.3. SYSTEM CLASS DIAGRAM

Figure 4.4: Class Diagram of Aspect-Based Sentiment Analysis in University and Student
System

The class diagram presented herein delineates the structural composition and interactions within
the aspect-based sentiment analysis (ABSA) system for course evaluation. Each class
encapsulates distinct entities and functionalities essential for the system's operation, facilitating a
comprehensive understanding of its design.

LI
4.4. AI PROCESS
In the context of our Aspect-Based Sentiment Analysis (ABSA) system, AI processes play a
pivotal role in enabling the system to analyze and understand the sentiments expressed in course
evaluation feedback. The AI processes employed in our system encompass various stages,
including data collection, data processing, model fine-tuning etc.

I. DATA COLLECTION

The dataset utilized in this study was obtained from the Hugging Face Datasets repository, a
comprehensive platform known for hosting a diverse range of datasets tailored for natural
language processing (NLP) tasks. Specifically, the dataset "course-reviews" by kkotkar1 was
selected due to its relevance to the research objective of aspect-based sentiment analysis in the
context of online course evaluations.

The "course-reviews" dataset, in its raw form, comprises user-generated reviews of various
online courses, offering valuable insights into the sentiments expressed towards different aspects
of course content, structure, and delivery. It is important to note that this raw dataset has not been
preprocessed to extract specific aspects, and it includes labels that may not be pertinent to the
research objectives. The dataset, while rich in content, required preprocessing to extract relevant
information and filter out irrelevant labels. Initially, the dataset contained raw text reviews
without specific annotations or aspect labels. Therefore, the first step in the data collection
process involved accessing the raw dataset from the Hugging Face repository and extracting the
textual content of the reviews.

Furthermore, it is essential to highlight that Hugging Face is not only a repository for datasets
but also hosts powerful models that can be leveraged for various NLP tasks. While the primary
focus of this study was data collection, the availability of pre-trained models on Hugging Face
may offer opportunities for future research endeavors or model fine-tuning. The dataset obtained
from Hugging Face constitutes secondary data, as it was collected and curated by the dataset
creator(s) for a different purpose.

In summary, the data collection process involved accessing the raw "course-reviews" dataset
from the Hugging Face repository, which provided a rich source of user-generated reviews

LII
relevant to the research topic. Subsequent preprocessing steps were necessary to extract relevant
information and filter out irrelevant labels before further analysis could be conducted.

II. DATA PRE-PROCESSING

The foundation of any machine learning endeavor rests upon the quality and relevance of the
dataset utilized for model training and fine-tuning. In the context of this research on aspect-based
sentiment analysis within course reviews, meticulous attention was directed towards acquiring a
dataset that encapsulated the nuanced sentiments associated with various facets of course
delivery.

Figure 4.5: Image showing the pre-processed JSONL data for fine-tuning

The dataset employed in this study was sourced from the Hugging Face Datasets repository, a
renowned hub for machine learning datasets and models. Notably, the dataset contains
unstructured reviews, lacking division into specific aspects pertinent to aspect-based sentiment
analysis. Moreover, irrelevant labels and noise are present within the dataset, necessitating
careful preprocessing to enhance its quality and relevance. Furthermore, it is essential to adhere
to the guidelines outlined by OpenAI for fine-tuning, which mandate the adoption of the JSONL

LIII
format for dataset preparation. Accordingly, the dataset is structured to conform to this format,
facilitating seamless integration into the fine-tuning process. Each instance within the dataset
encapsulates a conversation-like format, comprising messages categorized by role (system, user,
assistant), with corresponding content tailored to meet the requirements of the fine-tuning
process.

To ensure compliance with OpenAI's regulations and optimize the dataset for fine-tuning, the
preparation process was guided by the platform's documentation on dataset formatting. This
meticulous approach guarantees the integrity and efficacy of the dataset for subsequent stages of
model training and evaluation.

III. FINE-TUNING OF MODEL

The process of fine-tuning the OpenAI’s GPT-3.5 model involves utilizing the code directly or
the Graphical User Interface (GUI) provided by OpenAI's platform. For this research the data
was fine-tuned via the Graphical User Interface (GUI).

Through this interface, users can select from a variety of base models (see Figure 4.3) such as
babbage-002, davinci-002, gpt-3.5-turbo-0125, gpt-3.5-turbo-0163, and gpt-3.5-turbo-1160. For
this research format, the gpt-3.5-turbo-1160 model was selected.

Once the model is chosen, users upload both the training data and a validation dataset. It's
noteworthy that the validation data follows the same format as the training data, adhering to the
JSONL format required by Open AI for fine-tuning

After the datasets are uploaded, users proceed to create a fine-tuning job directly through the
GUI. Upon initiating the fine-tuning job, OpenAI's platform automatically provides progress
updates, including training loss and validation loss metrics over time. These metrics enable users
to monitor the model's performance throughout the fine-tuning process. Additionally, users
receive email confirmations upon completion of the fine-tuning job.

LIV
Figure 4.6: Completed Fine-tune Status Update Provided By Open AI Containing Trained
Tokens, Number of Epochs and Training and Validation Loss

The initial fine-tuning job for this research trained 89,550 tokens over 3 epochs. The training loss
was recorded as 0.0000, with a validation loss of 0.2196.

Tokens refer to individual units of text, typically words or sub words that the model processes
during training. In this case, the model was trained on a dataset containing 89,550 tokens, which
means it learned from a substantial amount of textual data.

An epoch represents one complete pass through the entire training dataset during model training.
Training over multiple epochs allows the model to learn from the data iteratively. In this
scenario, the model was trained for 3 epochs, meaning it went through the entire dataset three
times during the training process.

LV
The training loss is a measure of how well the model is performing during training. It represents
the error or difference between the predicted outputs of the model and the actual targets in the
training data.

The validation loss is similar to the training loss but is calculated on a separate dataset called the
validation set, which the model has not seen during training. It serves as a measure of the model's
generalization performance. A validation loss of 0.2196 indicates the average error of the
model's predictions compared to the actual targets on the validation set. Lower validation loss
values suggest better generalization performance, although the interpretation may vary
depending on the specific context and the scale of the loss values.

When we train a machine learning model, its primary objective is not only to learn the patterns
present in the training data but also to generalize well to unseen data. In other words, the model
should be able to make accurate predictions or classifications on new, previously unseen
examples that it hasn't encountered during training. The ability of a model to perform well on
such unseen data is known as its generalization performance.

The validation loss serves as an important indicator of the model's generalization performance.
During training, the model adjusts its parameters (weights and biases) to minimize the training
loss, which measures how well it's fitting the training data. However, the training loss alone
doesn't provide a complete picture of the model's performance because the model could
potentially memorize the training data without truly learning the underlying patterns.

Therefore, when we interpret the validation loss of 0.2196 in the context of the initial fine-tuning
job, it indicates that the model is performing reasonably well on unseen data, although there
might still be room for improvement. Lower validation loss values generally suggest better
generalization performance, and optimizing the model further to reduce this loss could
potentially lead to even better performance on new, unseen data.

IV. DEPLOYMENT

Following the fine-tuning process, the fine-tuned model is maintained by OpenAI and accessible
for inference via an API (Application Programming Interface). This deployment method offers a
straightforward way to utilize the model's capabilities without the need for local hosting or

LVI
management. The fine-tuned model is deployed on OpenAI's servers and can be invoked through
API endpoints for inference. Users can submit input data to the API, and the model generates
responses or predictions based on the provided inputs. The API-based deployment enables
remote access to the fine-tuned model from any location with internet connectivity. This
accessibility facilitates integration with various applications and systems, regardless of the user's
geographical location.

Utilizing the model through an API grants users the flexibility to incorporate its functionality
into their existing software infrastructure seamlessly. This approach eliminates the need for users
to manage the underlying infrastructure or worry about resource provisioning. By leveraging
OpenAI's infrastructure, the fine-tuned model remains continuously available for inference
requests, ensuring reliable performance and minimal downtime. Users can access the API
endpoints at any time to utilize the model's capabilities as needed.

OpenAI's deployment infrastructure is designed to handle varying levels of demand, allowing the
fine-tuned model to scale automatically based on the volume of inference requests. This
scalability ensures consistent performance, even during periods of increased usage.

4.5. BACKEND PROCESS


In the development of the system integrating the ABSA fine-tuned model, the backend serves as
a critical component responsible for orchestrating communication between users, databases, and
the OpenAI inference API. Leveraging Node.js along with the Express framework, the backend
is structured to handle various functionalities efficiently.

I. ROUTING APIS

The backend is organized into three main routes: student, university, and utility. These routes
serve distinct purposes, such as managing user authentication, handling form submissions, and
facilitating utility functions.

- Student Route: This route manages operations related to student accounts, including
registration, authentication, and form submissions for course evaluations.

LVII
- University Route: Responsible for university-related operations such as registration,
authentication, and managing course feedback forms.

- Utility Route: Handles miscellaneous utility functions required by the system.

Middleware functions play a pivotal role in intercepting and processing incoming requests before
they reach the designated endpoints. In this system, middleware functions are implemented
primarily for authorization purposes, ensuring that only authenticated users can access protected
resources.

III. DATABASE INTERACTIONS

MongoDB, a popular NoSQL database, is employed as the backend's data store due to its
flexibility and scalability. The Mongoose ORM library facilitates seamless interaction between
the Node.js runtime and MongoDB, simplifying data modeling and manipulation.

Figure 4.7: student database schema

LVIII
Before interacting with the database, data is structured using predefined schemas to ensure
consistency and integrity. These schemas define the structure of documents stored in MongoDB,
enforcing validation rules and maintaining data integrity.

IV. INTERACTION WITH THE ABSA FINE-TUNED MODEL

The backend integrates with the ABSA fine-tuned model provided by OpenAI to perform aspect-
based sentiment analysis on course reviews submitted by students.

Figure 4.8: inference point interacting with the fine-tuned model

Utilizing the OpenAI library, the backend sends requests to the inference API with specific
parameters, including the role (system or user) and the model name obtained from OpenAI after
fine-tuning. Upon receiving responses from the model, the backend stores the extracted aspects
in the database for further analysis and visualization by educational institutions.

LIX
V. DEPLOYMENT

For deployment, the backend utilizes Vercel's serverless functions, offering scalability and ease
of deployment without the overhead of managing infrastructure. By leveraging serverless
architecture, the system can dynamically scale based on demand, ensuring optimal performance
under varying workloads. Additionally, the MongoDB database is deployed on Mongo Atlas,
providing a robust cloud-based solution for data storage and management.

4.6. FRONTEND PROCESS


4.6.1. TOOLS AND TECHNOLOGIES USED
I. React.js: Chosen for its component-based architecture, which facilitates modular development
and reusability of UI elements.

II. Axios: Axios is a widely used JavaScript library for making HTTP requests, both in the
browser and in Node.js environments. It is favored for its simplicity, ease of use, and powerful
features that make handling HTTP communications more efficient and straightforward, it was
used to communicate with the server side of the system.

III. react-chartjs-2: react-chartjs-2 was used to create interactive and visually appealing charts,
enabling users to analyze and interpret data effectively.

IV. Lottiefile: Animations were obtained from lottiefile, contributing to a visually engaging user
experience.

V. React-Toastify : React-Toastify was used to Manage error handling, warning messages, and
success messages in a visual appealing way, providing users with timely feedback and improving
overall user experience.

4.6.2. INTEGRATION WITH BACKEND


I. RESTful API: Communication with the backend server was established through RESTful API
endpoints, ensuring standardized and efficient data exchange.

LX
4.6.3. FUNCTIONALITIES
I. User authentication:

Implemented using `@auth0/auth0-react`, a library that simplifies authentication processes and


integrates seamlessly with React applications.

II. Form submission:

Enables users to submit course reviews from the student side and create forms from the
educational side, facilitating seamless data input and processing.

III. Form control:

Provides functionalities such as editing, deleting, and checking submitted reviews from the
university side, enhancing administrative control and data management.

4.6.4. USER INTERFACE DESIGN

Figure 4.9: university dashboard portal

The User Interface (UI) Adopts a monochrome design for the user interface, ensuring a clean and
minimalist aesthetic that enhances easy usability and readability.

LXI
4.6.5. DEPLOYMENT
The Deployed the frontend application on Vercel, a platform known for its simplicity and
scalability, ensuring reliable hosting and seamless deployment processes.

4.7. TESTING AND EVALUATION


Testing is a critical phase in the development lifecycle of the Aspect-Based Sentiment Analysis
(ABSA) system, ensuring that the system functions as intended and meets the specified
requirements. Testing involves validating the system's behavior, functionality, and performance
against predefined criteria.

4.7.1 MODEL EVALUATION


Throughout the fine-tuning process, OpenAI provided evaluation metrics such as training loss
and validation loss at different time steps.

LXII
LXIII
Figure 4.10: Evaluation metrics showing training loss and validation loss for fine-tuned model at
different steps

These metrics were used to monitor the model's convergence and performance, ensuring that it
achieved desirable accuracy and generalization on unseen data. A training loss of 0.0000
indicates that the model achieved perfect or near-perfect performance on the training data, with
minimal error.

In the context of this research, manual evaluation of the model stands as an indispensable
measure to ensure the thorough examination of its performance. However, before delving into
the manual evaluative process, it becomes imperative to establish a profound comprehension of
fundamental evaluation concepts. This preliminary undertaking not only bolsters the
methodological integrity of the research but also lays the groundwork for a nuanced and rigorous
assessment of the fine-tuned model.

1. True Positive (TP):

I. Definition: The number of correctly predicted positive instances.

II. Case Study Application: In our case study, a true positive would be when the model
correctly predicts a positive aspect (e.g., predicting "positive" for "Assessment Feedback and
Fairness").

2. False Positive (FP):

I. Definition: The number of incorrectly predicted positive instances.

II. Case Study Application: In our case study, a false positive would occur if the model predicts
a positive aspect when it's not actually positive (e.g., predicting "positive" for "Learning
Environment" which is actually negative).

3. True Negative (TN):

I. Definition: The number of correctly predicted negative instances.

LXIV
II. Case Study Application: A true negative would happen if the model predicts a negative
aspect correctly (e.g., predicting "negative" for "Instructional Effectiveness" when is actually
negative).

4. False Negative (FN):

- Definition: The number of incorrectly predicted negative instances.

- Case Study Application: A false negative would happen if the model predicts a negative
aspect when it's actually positive (e.g., predicting "negative" for "Instructional Effectiveness"
which is actually neutral).

5. Precision:

I. Definition: The ratio of correctly predicted positive observations to the total predicted
positive observations.

II. Formula:

- Case Study Application: Precision measures the accuracy of positive predictions made by the
model. For our case study, precision would indicate how many positive predictions were correct
out of all the positive predictions made by the model.

6. Recall (Sensitivity):

I. Definition: The ratio of correctly predicted positive observations to the all observations in
actual class.

II. Formula:

- Case Study Application: Recall measures the ability of the model to capture all the positive
instances. In our case study, recall would indicate how many positive aspects were correctly
identified out of all the positive aspects present in the data.

LXV
7. Accuracy:

I. Definition: Accuracy is a measure of the overall correctness of the model's predictions,


indicating the proportion of correctly classified instances among the total instances.

II. Formula:

Where:

True Positives (TP) are the correctly predicted positive instances.

True Negatives (TN) are the correctly predicted negative instances.

Total is the total number of instances (sum of TP, TN, false positives, and false negatives).

Note: Accuracy provides a general view of how well the model performs across all classes or
aspects. However, it may not be suitable for imbalanced datasets where one class is dominant, as
the model can achieve high accuracy by simply predicting the dominant class most of the time.

8. F1 Score:

I. Definition: The harmonic mean of precision and recall, providing a balance between the two
metrics.

II. Formula:

- Case Study Application: F1 score takes into account both precision and recall. It's useful
when you want to seek a balance between precision and recall. In our case study, the F1 score
would provide an overall measure of the model's performance in identifying both positive and
negative aspects accurately.

Consider the open-ended course review below

LXVI
“I thoroughly enjoyed the course on Introduction to Data Science. The instructor was
knowledgeable and engaging, making complex concepts easy to understand. The assignments
were challenging yet manageable, allowing me to apply what I learned. However, I felt that the
pace of the course was a bit fast at times, and there could have been more opportunities for
interactive discussions during lectures. Overall, it was a valuable learning experience, and I
would recommend it to others interested in the field.”

Now based on the review above the table below shows the actual aspects against the fine-tuned
model aspect.

Table 4.1: Aspect-Based Sentiment Analysis Results Comparison


Aspect Actual Sentiment Predicted Sentiment

Instructional Effectiveness Positive Positive

Learning Environment Positive Positive

Assessment Feedback & Fairness Neutral Neutral

Assessment Evaluation Methods Neutral Neutral

Lecture Quality Positive Positive

Instructor Availability Neutral Neutral

Classroom Interaction Neutral Neutral

After The Precision, Recall, Accuracy and F1 score has been calculated using the formulas
provided earlier the table below shows the values.

LXVII
Table 4.2: Evaluation Metrics for Aspect-Based Sentiment Analysis Model
Actual Predicted True False False True
Sentiment Sentiment Positive Positive Negative Negative F1
Aspect (TP) (FP) (FN) (TN) Precision Recall Accuracy Score
Instructional Positive Positive
Effectiveness 1 0 0 0 1.00 1.00 1.00 1.00
Learning Positive Positive
Environment 1 0 0 0 1.00 1.00 1.00 1.00
Assessment Neutral Neutral
Feedback &
Fairness
Assessment Neutral Neutral
Evaluation
Methods
Lecture Positive Positive
Quality 1 0 0 0 1.00 1.00 1.00 1.00
Instructor Neutral Neutral
Availability
Classroom Neutral Neutral
Interaction 1 0 0 0 1.00 1.00 1.00 1.00

Interpretation of Manual Model Evaluation Results

The evaluation of the sentiment analysis model's performance, as depicted in Table 4.2, provides
valuable insights into its effectiveness and reliability in analyzing course evaluation feedback.
The table presents metrics such as True Positive (TP), False Positive (FP), False Negative (FN),
True Negative (TN), Precision, Recall, Accuracy, and F1 Score for each aspect considered in the
model. For aspects where both the actual sentiment and predicted sentiment are available, such
as Instructional Effectiveness, Learning Environment, and Lecture Quality, the model
demonstrates exceptional performance. These aspects show perfect scores across all evaluation
metrics, including Precision, Recall, Accuracy, and F1 Score. This indicates that the model

LXVIII
accurately predicts sentiments for these aspects, achieving a balance between correctly
identifying positive sentiments and avoiding false positives and negatives. However, in cases
where the actual sentiment is neutral, such as Assessment Feedback & Fairness, Assessment
Evaluation Methods, Instructor Availability, and Classroom Interaction, there are no true
positives, false positives, or false negatives recorded. This absence of data in these fields is due
to the nature of neutral sentiments, which indicate that the review does not cover these aspects.
As a result, it is not meaningful to calculate precision, recall, accuracy, or F1 score for these
aspects because there are no predictions to evaluate against. In conclusion, while the sentiment
analysis model performs exceptionally well for aspects with identifiable sentiments, the absence
of data for neutral aspects underscores the importance of considering the coverage and
comprehensiveness of course evaluation feedback. It highlights the need for further refinement
and validation efforts to ensure that the model accurately captures sentiments across all relevant
aspects of course evaluations.

4.7.2 BACKEND TESTING: UNIT TESTING ON ASPECT EVALUATION


EXTRACTION FOR COURSE EVALUATION

Figure 4.11: terminal showing that the test was successful

LXIX
Table 4.3: Test Case Description and Procedure for getCourseEvaluationAspects

SECTION DETAILS

I. Test Description This test case examines the efficiency of the getCourseEvaluationAspects utility
function, which serves to analyze evaluation comments pertaining to course
experiences and extract pertinent evaluation aspects.
II. Test Objective The primary objective is to validate the functionality and reliability of the
getCourseEvaluationAspects utility function in accurately identifying and
extracting relevant evaluation aspects from evaluation comments.
III. Test Environment The test is conducted within the context of a software development project where
course evaluation feedback analysis is integral. The environment includes the
necessary infrastructure to execute the utility function and assess its performance.
IV. Test Setup The test imports the getCourseEvaluationAspects utility function from the
corresponding module. Timeout Configuration: The timeout interval for the test is
set to 30,000 milliseconds jasmine.DEFAULT_TIMEOUT_INTERVAL = 30000;
to accommodate potential delays in asynchronous operations.
V. Test Procedure I. Mock Evaluation Comment Generation - A synthetic evaluation comment, “The
lecture quality exceeded my expectations”, is generated to simulate typical
feedback provided by students regarding their course experiences.

II. Utility Function Invocation: - The getCourseEvaluationAspects function is


invoked with the mock evaluation comment as its input parameter. - The utility
function asynchronously processes the evaluation comment to identify and extract
relevant evaluation aspects.

III. Assertion and validation - The test verifies the integrity of the extracted
evaluation aspects by assessing whether it is defined, signifying successful
execution of the utility function. - Further assertions, based on the anticipated
behavior of the getCourseEvaluationAspects function, can be included to validate
specific aspects extracted from the evaluation comment.
VI. Expected Outcome The anticipated outcome entails the successful execution of the

LXX
getCourseEvaluationAspects function, culminating in the extraction of pertinent
evaluation aspects from the provided evaluation comment. - The existence of a
defined aspects variable post-execution confirms the effective functioning of the
utility, thereby fulfilling the test's objectives.

VII. Result Interpretation:

i. Test Success: The test is deemed successful if it concludes without encountering errors, and all
assertions validate the expected behavior of the `getCourseEvaluationAspects` utility function.

ii. Test Failure: Conversely, failure occurs if errors impede test completion or if assertions fail to
corroborate the anticipated behavior of the utility function. Such outcomes necessitate further
investigation and potential refinement of the utility's implementation.

Viii. Importance:

- This test serves as a critical component of the quality assurance process within the software
development lifecycle (agile), ensuring the reliability and accuracy of course evaluation feedback
analysis.

- By validating the efficacy of the getCourseEvaluationAspects utility function, this test


contributes to enhancing the overall robustness and functionality of the course evaluation system,
thereby fostering improved decision-making based on insightful feedback analysis.

LXXI
4.7.3 FRONTEND TESTING
I. Testing:

No specific frontend testing strategy was done via react testing tools, however, a Chrome
Lighthouse test was conducted to assess performance, accessibility, best practices, SEO, and
PWA compliance.

Lighthouse Test Results:

Figure 4.12: chrome lighthouse result on the ABSA for course evaluation platform

Conducted a Chrome Lighthouse test to evaluate various aspects of the frontend, including
performance, accessibility, best practices, SEO, and PWA compliance.

Results indicate areas of improvement and adherence to web development standards, ensuring
optimal performance and user accessibility.

LXXII
4.8 DISCUSSIONS OF FINDINGS

RESEARCH QUESTION I: HOW CAN ASPECT-BASED SENTIMENT ANALYSIS


(ABSA) BE EFFECTIVELY APPLIED TO ENHANCE THE ANALYSIS OF COURSE
EVALUATIONS AND REVEAL MULTIFACETED STUDENT FEEDBACK?

The implementation of Aspect-based Sentiment Analysis (ABSA) in the analysis of course evaluations
has proven to be highly effective in extracting nuanced insights from student feedback. By fine-tuning a
Large Language Model (LLM) specifically on course reviews and their aspects, we have developed a
system capable of detecting multifaceted insights in student course reviews. This approach allows for a
granular analysis of feedback, identifying specific aspects of the course experience and associating
sentiment polarities with these aspects. This enables educational institutions to pinpoint exact areas of
strength and weakness, facilitating targeted improvements in course delivery and design.

RESEARCH QUESTION II: WHAT ARE THE SPECIFIC ASPECTS OF STUDENT


FEEDBACK IN COURSE EVALUATIONS, AND HOW DO SENTIMENT POLARITIES
ASSOCIATED WITH THESE ASPECTS PROVIDE INSIGHTS INTO THE STRENGTHS AND
WEAKNESSES OF COURSES?

Through the ABSA methodology, we identified several key aspects in student feedback that are critical to
understanding the overall course experience. These aspects include:

Instructional Effectiveness: Refers to the ability of the instructor to effectively deliver course content.

Learning Environment: Encompasses the physical and psychological conditions of the learning space.

Assessments and Feedback: Relates to the quality and timeliness of assessments and feedback provided to
students.

Assessment Evaluation Methods: Focuses on the appropriateness and fairness of evaluation methods.

Lecture Quality: Evaluates the clarity, engagement, and informativeness of lectures.

Instructor Availability: Measures how accessible and responsive instructors are to students.

Classroom Interaction: Assesses the level and quality of interaction between students and instructors, as
well as among students.

LXXIII
By analyzing the sentiment polarities associated with these aspects, we gain valuable insights into the
strengths and weaknesses of the courses. For instance, positive sentiment in "Instructional Effectiveness"
and "Lecture Quality" indicates effective teaching practices and engaging content delivery, while negative
sentiment in "Assessment Evaluation Methods" and "Instructor Availability" highlights areas needing
improvement. This detailed feedback is crucial for educational institutions aiming to enhance their
teaching strategies and overall course quality.

RESEARCH QUESTION III: HOW DO EDUCATIONAL INSTITUTIONS GET THE


STUDENTS TO SUBMIT EVALUATIONS FOR A COURSE, HOW DO STUDENTS SUBMIT
THE EVALUATIONS, AND HOW DO THE ASPECTS FROM THE EVALUATIONS GET
CONVEYED OVER TO THE EDUCATIONAL INSTITUTIONS?

A dedicated platform was developed to facilitate the submission and analysis of course evaluations. This
platform serves as a bridge between students and educational institutions, streamlining the evaluation
process and enhancing the interaction with ABSA for course evaluations.

LXXIV
CHAPTER FIVE

SUMMARY AND CONCLUSION


5.1 SUMMARY
This research endeavors to develop an Aspect-Based Sentiment Analysis (ABSA) system for
analyzing course evaluation feedback. The system aims to provide insights into various aspects
of course delivery, such as instructional effectiveness, learning environment, and lecture quality,
by analyzing sentiments expressed in student reviews. Chapter Four detailed the design and
modeling of the proposed system, including the use case diagram, activity flow diagram, and
class diagram, highlighting the system's architecture and functionalities.

The proposed ABSA system incorporates both functional and non-functional requirements to
ensure scalability, security, accuracy, usability, performance, and continuous availability.
Testing procedures, such as unit tests on aspect evaluation extraction, are employed to validate
the system's behavior and functionality.

5.2 CONTRIBUTION OF RESEARCH AND CONCLUSION


This research makes several significant contributions to the field of educational feedback
analysis:

1. Development of ABSA System: The research proposes a novel ABSA system tailored
specifically for analyzing course evaluation feedback. By leveraging sentiment analysis
techniques, the system provides detailed insights into various aspects of course delivery,
facilitating informed decision-making by educational institutions.

2. Comprehensive Design and Modeling: Through the use of use case diagrams, activity flow
diagrams, and class diagrams, the research offers a comprehensive overview of the ABSA
system's architecture and functionalities. This structured approach enhances understanding and
facilitates effective implementation of the system.

3. Testing and Validation: The research emphasizes the importance of testing and validation in
ensuring the reliability and accuracy of the ABSA system. By conducting unit tests and defining
clear testing procedures, the research ensures the robustness of the system's functionality.

LXXV
In conclusion, the proposed ABSA system represents a valuable tool for educational institutions
seeking to gain deeper insights into student perceptions of course delivery. By systematically
analyzing course evaluation feedback, the system empowers institutions to identify areas of
improvement and enhance the overall quality of education.

5.3 RECOMMENDATIONS
Based on the findings of this research, the following recommendations are provided:

1. Implementation and Deployment: Educational institutions are encouraged to implement and


deploy the ABSA system to analyze course evaluation feedback systematically. By integrating
the system into existing feedback mechanisms, institutions can enhance their ability to address
student concerns and improve course delivery.

2. Continuous Monitoring and Evaluation: It is recommended that institutions continuously


monitor and evaluate the performance of the ABSA system to ensure its effectiveness in
providing actionable insights. Regular feedback from users should be solicited to identify areas
for improvement and refinement.

3. Collaboration and Knowledge Sharing: Institutions are encouraged to collaborate and share
best practices in the implementation and utilization of ABSA systems. By fostering a culture of
knowledge sharing, institutions can collectively advance the field of educational feedback
analysis.

5.4 AREA FOR FURTHER RESEARCH


While this research lays a strong foundation for the development and implementation of ABSA
systems in educational settings, there are several areas for further research:

1. Enhanced Sentiment Analysis Techniques: Future research could explore advanced sentiment
analysis techniques, such as deep learning and natural language processing, to improve the
accuracy and granularity of sentiment analysis in course evaluation feedback. Additionally,
research could focus on incorporating sentiment analysis for languages other than English, such
as Spanish, French, or Mandarin, to broaden the applicability of ABSA systems across diverse
linguistic contexts.

LXXVI
2. Integration with Learning Analytics: There is potential to integrate ABSA systems with
learning analytics platforms to gain deeper insights into the relationship between student
feedback and academic performance. This integration could facilitate personalized interventions
and support mechanisms for students. Furthermore, exploring how ABSA insights correlate with
student engagement metrics and learning outcomes could provide valuable insights for
instructional design and curriculum development.

3. Cross-Institutional Studies: Comparative studies across multiple educational institutions could


provide valuable insights into the effectiveness of ABSA systems in different contexts. By
examining variations in student perceptions and feedback patterns across institutions and cultural
settings, researchers can identify factors influencing course satisfaction and engagement.
Moreover, conducting cross-linguistic studies to compare the performance of ABSA models
trained on different languages could contribute to understanding cultural nuances in feedback
interpretation and sentiment analysis.

In summary, further research in these areas has the potential to advance the field of educational
feedback analysis and contribute to the ongoing efforts to enhance the quality of education
delivery across diverse linguistic and cultural contexts.

LXXVII
REFERENCES

Ajmal, F., Durrani, R., & Mohammad, N. (2021). Teaching effectiveness: A university students'
perspective. PJER, 4(4).

Al-Saqqa, S., Sawalha, S., & AbdelNabi, H. (2020, July 10). Agile is a wide umbrella of
software development beliefs. Agile Software Development: Methodologies and Trends.
Retrieved from https://doi.org/10.3991/ijim.v14i11.13269.

Ameer, A. M., & Hashim, U. R. (2020). The effectiveness of course evaluation on improving
teaching and learning quality. International Journal of Evaluation and Research in Education
(IJERE), 9(2), 432-439. doi:10.11591/ijere.v9i2.20579. Retrieved from
https://ijere.iaescore.com/index.php/IJERE/article/view/20579.

Bolukbasi, T., Chang, K. W., Zheng, J. Y., Li, Z., & Peng, Y. (2017, July). Quantifying the
language understanding ability of LSTMs for natural language processing tasks. Proceedings of
the Association for Computational Linguistics.

D’aniello, G., Gaeta, M., & La Rocca, I. (2022). KnowMIS-ABSA: An overview and a reference
model for applications of sentiment analysis and aspect-based sentiment analysis. Artificial
Intelligence Review, 55, 5543-5574.

Dai, A., Hu, X. H., Nie, J., & Chen, J. (2022). Learning from word semantics to sentence syntax
by graph convolutional networks for aspect-based sentiment analysis. International Journal of
Data Science and Analytics, 14, 17-26.

Fei, H., Chua, T., Li, C., Ji, D., Zhang, M., & Ren, Y. (2022). On the robustness of aspect-based
sentiment analysis: Rethinking model, data, and training. ACM Transactions on Information
Systems, 41(1), 1-32.

Fei, H., Li, F., Li, C., Wu, S., Li, J., & Ji, D. (2022). Inheriting the wisdom of predecessors: A
multiplex cascade framework for unified aspect-based sentiment analysis. International Joint
Conference on Artificial Intelligence.

He, K., Mao, R., Gong, T., Li, C., & Cambria, E. (2022). Meta-based self-training and re-
weighting for aspect-based sentiment analysis. IEEE Transactions on Affective Computing.

LXXVIII
He, R., Lee, W. S., Ng, H. T., & Dahlmeier, D. (2019). An interactive multi-task learning
network for end-to-end aspect-based sentiment analysis. Annual Meeting of the Association for
Computational Linguistics.

Hugging Face Datasets Repository. (n.d.). Course-reviews dataset by kkotkar1. Retrieved from
https://huggingface.co/datasets/kkotkar1/course-reviews.

Lalata, J. P., Gerardo, B., & Medina, R. M. (2019). A sentiment analysis model for faculty
comment evaluation using ensemble machine learning algorithms. Proceedings of the 2019
International Conference on Big Data Engineering.

Liang, S., Wei, W., Mao, X., Wang, F., & He, Z. (2022). BiSyn-GAT+: Bi-Syntax aware graph
attention network for aspect-based sentiment analysis. Findings of the Association for
Computational Linguistics.

Lundqvist, K., Liyanagunawardena, T. R., & Starkey, L. (2020). Evaluation of student feedback
within a MOOC using sentiment analysis and target groups. The International Review of
Research in Open and Distributed Learning.

Mao, Y., Shen, Y., Yu, C., & Cai, L. (2021). A joint training dual-MRC framework for aspect-
based sentiment analysis. AAAI Conference on Artificial Intelligence.

Onan, A. (2020). Sentiment analysis on massive open online course evaluations: A text mining
and deep learning approach. Computer Applications in Engineering Education, 29, 572-589.

OpenAI. (n.d.). Preparing your dataset. Retrieved from


https://platform.openai.com/docs/guides/fine-tuning/preparing-your-dataset.

Pontiki, M., Galanis, D., Papageorgiou, H., Manandhar, S., Androutsopoulos, I., Bel, N., &
Eryiğit, G. (2016). SemEval-2016 Task 5: Aspect-based sentiment analysis. International
Workshop on Semantic Evaluation.

Pontiki, M., Galanis, D., Pavlopoulos, J., Papageorgiou, H., Androutsopoulos, I., & Manandhar,
S. (2014). SemEval-2014 Task 4: Aspect-based sentiment analysis. International Workshop on
Semantic Evaluation.

LXXIX
Pu, X., Yan, G., Yu, C., Mi, X., & Yu, C. (2021). Sentiment analysis of online course evaluation
based on a new ensemble deep learning model: Evidence from Chinese. Proceedings of the
Association for Computational Linguistics.

Sun, C., Huang, L., & Qiu, X. (2019). Utilizing BERT for aspect-based sentiment analysis via
constructing auxiliary sentence. ArXiv, abs/1903.09588.

Suprianto, Fadlan, M., Muhammad, Amaliah, Y., & Mussallimah. (2020). Retrieval information
using generalized vector space models and sentiment analysis using Naïve Bayes classifier for
evaluation of lecturers by students. 2020 Fifth International Conference on Informatics and
Computing (ICIC), 1-7.

Wang, K., Shen, W., Yang, Y., Quan, X., & Wang, R. (2020). Relational graph attention network
for aspect-based sentiment analysis. Annual Meeting of the Association for Computational
Linguistics.

Wu, Y., Ming, Z., Allen, J. K., & Mistree, F. (2022). Evaluation of students’ learning through
reflection on doing based on sentiment analysis. 19th International Conference on Design
Education (DEC).

Xu, H., Liu, B., Shu, L., & Yu, P. S. (2019). BERT post-training for review reading
comprehension and aspect-based sentiment analysis. North American Chapter of the Association
for Computational Linguistics.

Yan, H., Dai, J., Ji, T., Qiu, X., & Zhang, Z. (2021). A unified generative framework for aspect-
based sentiment analysis. ArXiv, abs/2106.04300.

Zhang, K., Zhang, K., Zhang, M., Zhao, H., Liu, Q., Wu, W. Y., & Chen, E. (2022).
Incorporating dynamic semantics into pre-trained language model for aspect-based sentiment
analysis. Findings of the Association for Computational Linguistics.

Zhang, W., Li, X., Deng, Y., Bing, L., & Lam, W. (2022). A survey on aspect-based sentiment
analysis: Tasks, methods, and challenges. ArXiv, abs/2203.01054.

Zhang, Z., Zhou, Z., & Wang, Y. (2022). SSEGCN: Syntactic and semantic enhanced graph
convolutional network for aspect-based sentiment analysis. North American Chapter of the
Association for Computational Linguistics.

LXXX
Zhao, J., Yang, X., Qiao, Q., & Chen, L. (2020). Sentiment analysis of course evaluation data
based on SVM model. 2020 IEEE International Conference on Progress in Informatics and
Computing (PIC), 375-379.

Zhong, Q., Ding, L., Liu, J., Du, B., Jin, H., & Tao, D. (2022). Knowledge graph augmented
network towards multiview representation learning for aspect-based sentiment analysis. ArXiv,
abs/2201.04831.

LXXXI
APPENDIX

APPENDIX 1
ETHICAL CONSIDERATIONS IN USING ASPECT-BASED SENTIMENT ANALYSIS FOR
COURSE EVALUATION

I. Privacy and Confidentiality


1. Balancing Anonymity and Integrity: While ensuring feedback integrity by verifying student
identities, protect their privacy using anonymized identifiers.
2. Data Security: Implement robust data protection measures such as encryption and restricted
access to safeguard feedback data.

II. Informed Consent


1. Transparency: Clearly inform students about the purpose of the sentiment analysis and how
their feedback will be used.
2. Voluntary Participation: Ensure that participation is voluntary and that students understand
they can opt out without any negative consequences.

III. Bias and Fairness


1. Algorithmic Bias: Regularly evaluate and adjust sentiment analysis algorithms to prevent bias.
2. Inclusive Representation: Strive to include diverse student feedback to ensure comprehensive
analysis.

IV. Interpretation and Use of Results


1. Contextual Interpretation: Interpret results within the broader context of course evaluation to
avoid misinterpretation.
2. Constructive Use: Use the insights to improve course quality and address genuine student
concerns.

V. Ethical Research Practices


1. Compliance: Follow ethical guidelines and institutional review board (IRB) requirements.

LXXXII
2. Continuous Review: Regularly update ethical practices to address new challenges and
maintain high standards.

LXXXIII
APPENDIX 2
USER INTERFACE DESIGN

LXXXIV
LXXXV
LXXXVI
APPENDIX 3 SOURCE CODE

LXXXVII
LXXXVIII

You might also like