A Hybrid CNN-LSTM Model for Opinion Mining and Classiﬁcation of

Course Reviews

Hatem Majouri

, Olfa Gaddour

and Yessine Hadj Kacem

CES Laboratory, National Engineering, School of Sfax, University of Sfax, Tunisia

Keywords:

Deep-Learning, e-Learning Platforms, CNN-LSTM, Online Course Reviews, User Feedback Analysis.

Abstract:

Automatic analysis of online course reviews is a critical task that has garnered signiﬁcant interest, particularly

for improving the quality of e-learning platforms. The challenge lies in accurately classifying user feedback

in order to generate actionable insights for educators and learners. In this work, we investigate the effec-

tiveness of a hybrid CNN-LSTM model compared to several state-of-the-art deep learning models, including

BERT, LSTM, GRU, and CNN, for analyzing reviews collected from the FutureLearn platform. Our experi-

ments demonstrate that the proposed model achieves superior performance in classifying user reviews, with

an accuracy of 0.95. These results highlight the potential of advanced deep learning techniques in extracting

meaningful insights from user feedback, offering valuable guidance for course developers and learners.

1 INTRODUCTION

The rise of online education has transformed the

landscape of learning, offering unprecedented access

to courses across a wide range of subjects. E-learning

platforms such as FutureLearn have gained popularity

by enabling learners to acquire new skills and knowl-

edge from the comfort of their homes. However,

with the increasing number of courses available,

evaluating and improving the quality of online

education has become crucial. User reviews provide

valuable insights into learner satisfaction, course

effectiveness, and areas for improvement. Effective

analysis of these reviews is vital for enhancing the

learning experience and helping both learners and

course providers make informed decisions.

Review analysis—the process of automatically

detecting and classifying opinions expressed in

text—has emerged as a powerful tool for understand-

ing user feedback. Traditional methods of review

analysis often struggle with the nuances of language,

making it challenging to accurately capture user

opinions. However, recent advances in deep learning

have led to the development of sophisticated models

capable of addressing these limitations. Notable

models, such as BERT (Bidirectional Encoder Repre-

https://orcid.org/0009-0002-6629-8527

https://orcid.org/0000-0002-2693-2055

https://orcid.org/0000-0002-5757-6516

sentations from Transformers) (Devlin et al., 2019),

LSTM (Long Short-Term Memory) (Siami-Namini

et al., 2019), GRU (Gated Recurrent Unit) (She

and Jia, 2021), and CNN (Convolutional Neural

Network) (LeCun et al., 1998), have shown great

promise in text classiﬁcation tasks, enabling more

accurate review analysis.

In this paper, we explore the application of state-

of-the-art deep learning techniques to the task of re-

view analysis on a dataset of online course reviews

from the FutureLearn platform. The dataset was col-

lected using web scraping and structured to include

features such as course names, student reviews, and

ratings. Our study is distinctive in its comprehensive

evaluation of multiple deep learning models, includ-

ing BERT, LSTM, GRU, and CNN, for classifying

user reviews in the context of e-learning. By lever-

aging these advanced techniques, we aim to uncover

deeper insights into user feedback, offering a more

nuanced understanding of learner experiences. This

work not only contributes to the growing body of re-

search on feedback analysis in online education but

also provides practical implications for enhancing the

quality of e-learning platforms.

The remainder of this paper is structured as fol-

lows. Section 2 reviews related work on review anal-

ysis and the application of deep learning techniques

in the context of e-learning. Section 3 details the

methodology used for data collection, including the

web scraping process, dataset structuring, and de-

734

Majouri, H., Gaddour, O. and Kacem, Y. H.

A Hybrid CNN-LSTM Model for Opinion Mining and Classiﬁcation of Course Reviews.

DOI: 10.5220/0013177900003890

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

In Proceedings of the 17th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2025) - Volume 3, pages 734-741

ISBN: 978-989-758-737-5; ISSN: 2184-433X

scriptions of the deep learning models implemented

for feedback classiﬁcation. Section 4 presents the ex-

perimental results, highlighting the performance of

each model. Finally, Section 5 concludes the paper

with a summary of key contributions and suggestions

for future research.

2 RELATED WORK

The rise of e-learning platforms has spurred research

into user feedback analysis to enhance course quality.

This section examines classiﬁcation methods for stu-

dent reviews, focusing on datasets and methodologies

to identify trends and outline future research direc-

tions.

In (Onan, 2020), deep-learning techniques were

used to analyze Student Evaluations of Teaching

(SET) for assessing teaching effectiveness and guid-

ing administrative decisions. The study employed a

recurrent neural network (RNN) enhanced with an

attention mechanism and GloVe word embeddings,

showcasing the model’s capabilities. However, the

dataset, not designed for educational or sentiment

analysis, led to overﬁtting due to its small size rela-

tive to feature complexity.

In (Kastrati et al., 2020b), an aspect-based opin-

ion mining model improved feedback analysis in on-

line courses. Despite its potential, reliance on a single

dataset raised scalability concerns, and the model’s

architecture failed to capture semantic relationships

effectively due to limited CNN dimensions and sim-

plistic layers.

The authors in (Chakravarthy et al., 2021) empha-

sized the importance of qualitative feedback in online

education, which is often overshadowed by quantita-

tive data. They developed an opinion-mining frame-

work using NLP and machine learning to classify stu-

dent feedback from a Coursera course. However, their

ﬁndings were constrained by the dataset, and auto-

mated methods overlooked nuanced opinions.

In (Onan, 2021), advanced machine-learning

techniques, including ensemble learning and deep-

learning, were applied to MOOC reviews. The study

evaluated text representation and word-embedding

schemes on a dataset of 66,000 reviews but faced

challenges with model interpretability and limited

generalizability due to its reliance on a large dataset.

Research in (Mrhar et al., 2021) compared CNN,

LSTM, and CNN-LSTM models for sentiment analy-

sis in MOOCs. A key limitation was the reliance on

manually labeled datasets, which introduced subjec-

tivity and scalability issues for larger or more diverse

datasets.

In (Koufakou, 2023), deep-learning models like

CNN, BERT, RoBERTa, and XLNet were used for

sentiment analysis and topic classiﬁcation of student

feedback. While the study offered insights into model

optimization, it lacked exploration of model inter-

pretability and employed unbalanced datasets, skew-

ing results and impairing performance on smaller cat-

egories.

Other works, such as (El-Halees, 2011; Cabada

et al., 2018; Kastrati et al., 2020a; Yan et al., 2021;

Edalati et al., 2022; Shaik et al., 2023), illustrate the

effectiveness of machine learning and deep-learning

in course feedback analysis. Despite their contribu-

tions to enhancing the educational experience, these

studies often face challenges in scalability, model in-

terpretability, and adaptability across platforms.

3 PROPOSED APPROACH

This section presents the proposed approach, detail-

ing a reliable system for data collection and model

generalization across diverse online courses and dis-

ciplines. It describes the dataset size, preprocessing

techniques, and data augmentation to address class

imbalance. Dropout and L2 regularization are ap-

plied to mitigate overﬁtting. The hybrid CNN-LSTM

model leverages CNN to capture local features and

context, while LSTM preserves sequential relation-

ships, enabling comprehensive data analysis. Perfor-

mance is evaluated using metrics such as accuracy,

recall, and AUC, providing a robust assessment. This

approach represents a signiﬁcant improvement over

previous methods by combining advanced deep learn-

ing techniques with rigorous evaluation processes.

3.1 Proposed System Architecture

This subsection describes the proposed sentiment

analysis architecture for e-learning platform reviews.

As illustrated in Figure 1, the process starts with raw

data collection and cleaning to ensure quality. The

text is then tokenized or vectorized for model train-

ing, with the data split into training and testing sets

for evaluation. Optimized deep-learning models are

used to classify the reviews, improving accuracy and

reliability in sentiment detection.

3.2 Data Collection

Data collection includes key steps such as identifying

sources, web scraping, structuring the dataset, and de-

termining its size, all essential for ensuring data accu-

racy and usability.

A Hybrid CNN-LSTM Model for Opinion Mining and Classiﬁcation of Course Reviews

735

Figure 1: Overview of the Data Processing and Classiﬁca-

tion Pipeline.

3.2.1 Data Source

The data was collected from the FutureLearn plat-

form (MAGNONI and PLUTINO, 2018), an online

learning environment offering a wide range of courses

from universities and institutions worldwide. Future-

Learn enables students to engage in courses featuring

interactive elements such as quizzes and discussions.

Our focus was on compiling and analyzing student

evaluations of various courses, extracting feedback on

experiences, content, instructor quality, and overall

satisfaction. The objective was to gain insights into

the quality and effectiveness of the courses offered by

FutureLearn.

3.2.2 Web-Scraping

We used web scraping to collect data for our on-

line course review dataset, efﬁciently extracting re-

views and insights from various platforms (Khder,

2021). Python scripts were employed to analyze web-

site structures, with the Beautiful Soup library parsing

HTML and identifying relevant tags for extraction.

The gathered data was stored in CSV or XLSX ﬁles,

creating a comprehensive dataset of user reviews and

ratings.

3.2.3 Data Structure

The structure of a data set varies depending on the

type of data and the intended use, usually consisting

of rows and columns where each row represents an in-

dividual data point and each column denotes a speciﬁc

attribute. In this context, the dataset captures informa-

tion about online courses, student reviews, and corre-

sponding ratings, facilitating effective data processing

and analysis. Key features include course name as a

string, reviews as a string, and rate as a ﬂoating, sup-

porting comprehensive analysis of course feedback

and ratings.

Table 1: Dataset Size by Class (Original vs. After Clean-

ing).

Class Original Num-

ber of Samples

Number of

Samples After

Cleaning

1 19,271 18,978

0 7,932 5,228

Total 27,203 24,206

3.3 Data Labeling

Sentiment classiﬁcation can be categorized as either

binary, which involves the classiﬁcation of reviews

into positive or negative categories, or multi-class,

which encompasses labels such as strong positive,

positive, neutral, negative, and strong negative. The

application of binary classiﬁcation is prevalent in the

ﬁeld of sentiment analysis research (Tripathy et al.,

2016). Furthermore, in (Guru and Bajnaid, 2023),

the dataset underwent a relabeling process to facil-

itate binary sentiment classiﬁcation through the uti-

lization of TextBlob. Reviews exhibiting a positive

polarity were designated as positive, whereas those

demonstrating zero or negative polarity were classi-

ﬁed as negative. This methodological simpliﬁcation

aimed to enhance the differentiation between positive

and negative sentiments. The present study employed

web-scraping techniques to amass a total of 30,121 re-

views regarding online courses, which were initially

classiﬁed into ﬁve distinct categories according to a

rating scale ranging from 1.0 to 5.0. Due to a pro-

nounced deﬁciency of data within classes 1, 2, and

3, our analysis concentrated on binary classiﬁcation

by designating 1 and 2-star evaluations as negative

(0) and 3, 4, and 5-star evaluations as positive (1).

The transformation of a multi-class problem into a bi-

nary classiﬁcation framework is a prevalent technique

in sentiment analysis, which streamlines the task to

emphasize the dichotomy of positive versus negative

sentiments. As highlighted in (Pang and Lee, 2008),

this methodology adeptly captures critical differentia-

tions in opinionated text and has the potential to yield

more resilient models, particularly in contexts char-

acterized by imbalanced or sparse datasets. Follow-

ing the reclassiﬁcation process, we discarded empty

rows and duplicates, thereby enhancing the dataset to

accurately reﬂect the distribution of each binary cate-

gory. The conclusive speciﬁcations of the dataset are

presented in TABLE 1.

3.4 Data Preprocessing

The initial phase in the analysis of reviews involves

the meticulous preparation of textual data through

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

736

processes of cleansing and reﬁnement. A signiﬁcant

proportion of the unprocessed reviews obtained from

FutureLearn exhibit extraneous elements and absent

information, thereby necessitating a thorough prepro-

cessing to facilitate effective analytical procedures; to

partition the dataset into training and testing subsets,

an allocation of 80% is designated for training pur-

poses while 20% is reserved for testing.

3.4.1 Data Cleaning

Data cleansing is a critical process for enhancing the

quality of datasets by rectifying inconsistencies, er-

rors, and absent values. Fundamental activities en-

compass label encoding to transform target variables

into binary classiﬁcations, addressing missing and du-

plicate entries, and ensuring uniformity in data types.

Text preprocessing procedures involve the elimina-

tion of special characters, HTML elements, and ex-

traneous spaces, the expansion of contractions, to-

kenization, lowercasing, and lemmatization of lex-

emes. Stopwords are excluded, while signiﬁcant

negations such as ”no” and ”not” are preserved in or-

der to maintain the contextual integrity of sentiment.

These methodologies guarantee that the data is ade-

quately prepared for the training of models.

3.4.2 Data Transformation

Data transformation for text classiﬁcation varies by

model. LSTM, GRU, and CNN use tokenization

(splitting text into words or subwords), while BERT

relies on its specialized subword tokenizer. For

LSTM, GRU, and CNN, text normalization (e.g.,

lowercasing and punctuation removal) is applied,

whereas BERT handles this internally via its tok-

enizer. Padding and truncation ensure consistent

sequence lengths, performed manually for LSTM,

GRU, and CNN, but automated by BERT. Tokens are

converted into numerical formats using embeddings

or one-hot encoding for LSTM, GRU, and CNN, and

contextual embeddings for BERT. Categorical labels

are encoded numerically across all models.

3.5 CNN-LSTM Model Architecture

Our proposed hybrid CNN-LSTM architecture,

shown in Figure 2, is designed for text data analy-

sis and classiﬁcation. Input sentences are represented

as a matrix of size N × K, where N is the number of

sentences and K the number of features (e.g., word

embeddings). A convolutional layer processes this

matrix, applying ﬁlters to extract local features and

capture essential patterns and context.

The convolutional output is passed through a max-

pooling layer, which reduces dimensionality by se-

lecting maximum values from each region, retaining

critical features while reducing computational cost.

The pooled output is then fed into an LSTM layer,

which captures dependencies and contextual informa-

tion across sequences.

Subsequently, the features are processed by one

or more fully connected (dense) layers for high-level

reasoning. A dropout layer follows to prevent over-

ﬁtting by randomly deactivating neurons during train-

ing, enhancing generalization. The ﬁnal output is gen-

erated through a sigmoid activation function, produc-

ing a probability between 0 and 1, representing the

likelihood of each class. This output classiﬁes sen-

tences into two categories: Class 0 and Class 1.

By combining convolutional layers for feature ex-

traction with LSTM layers for sequence processing,

followed by dense and dropout layers, this architec-

ture achieves robust classiﬁcation performance.

4 EXPERIMENTAL RESULTS

AND DISCUSSION

4.1 Evaluation Metrics

To evaluate our algorithm’s performance, we use sev-

eral metrics. Accuracy, calculated as the proportion

of correctly classiﬁed instances among all instances,

is deﬁned as:

Accuracy =

T P + T N

T P + T N + FP + FN

where T P, TN, FP, and FN represent true posi-

tives, true negatives, false positives, and false nega-

tives, respectively. Precision indicates the proportion

of true positive predictions among all positive predic-

tions made by the model and is given by:

Precision =

T P

T P + FP

Recall represents the proportion of true positive pre-

dictions among all actual positive instances and is cal-

culated as:

Recall =

T P

T P + FN

The F1 Score, which is the harmonic mean of preci-

sion and recall, provides a single metric that balances

both aspects and is deﬁned as:

F1 Score =

2 × Precision × Recall

Precision + Recall

A Hybrid CNN-LSTM Model for Opinion Mining and Classiﬁcation of Course Reviews

737

Figure 2: Proposed Hybrid CNN-LSTM Model Architecture.

4.2 Performance Visualization

Key visualizations for assessing model performance

include the accuracy curve, which tracks changes in

accuracy over epochs, and the loss curve, which high-

lights error reduction and convergence. The confu-

sion matrix displays the distribution of true positives,

false positives, and false negatives across classes,

while the ROC curve demonstrates the trade-off be-

tween true and false positive rates at various thresh-

olds.

4.3 Experimental Scenario

The dataset was scraped, cleaned, tokenized, and split

into 80% training and 20% testing, with augmentation

applied only to the training data. Tokenized words

were vectorized using GloVe embeddings to capture

semantic associations. The model combines CNN

and LSTM layers for feature extraction and sequence

analysis. A non-trainable GloVe embedding layer was

followed by CNN layers with kernels, and ﬁnally,

an LSTM layer with regularization to mitigate over-

ﬁtting. Dropout layers further enhanced generaliza-

tion, while dense layers integrated features for classi-

ﬁcation. The model was trained using binary cross-

entropy loss and the Adam optimizer for 50 epochs

with a batch size of 32, tracking performance on the

test set.

The CNN architecture includes two Conv1D lay-

ers with 128 ﬁlters and a kernel size of 5, each fol-

lowed by MaxPooling1D layers with a pool size of

2. An additional Conv1D layer with 64 ﬁlters and

a kernel size of 3, followed by another MaxPool-

ing1D layer, is included. All CNN layers use dropout

with a rate of 0.3%. The dense layers after the CNN

have 32 and 16 units with ReLU activation and L2

regularization applied to both kernel and bias. The

LSTM layer matches the embedding dimension, with

return

equences = False to output the hidden state.

L2 regularization and a 0.3 dropout rate are also ap-

plied to the LSTM layer.

4.4 Results Without Data Augmentation

In the results section, we analyze the raw data pro-

cessed without augmentation, evaluating the perfor-

mance of GRU, LSTM, CNN, and BERT models.

Each model is assessed to highlight its ability to han-

dle the original data, along with its relative advantages

and limitations.

TABLE 2 presents the performance results of

these models applied to our dataset without any aug-

mentation.

Starting with BERT, the model shows a precision

of 56% for class 0 and 94% for class 1. The recall

values are 82% for both classes, and the F1 scores

are 67% for class 0 and 88% for class 1. The overall

accuracy of BERT is 82

Next, the LSTM model achieves a precision of

79% for class 0 and 88% for class 1. The recall for

class 0 is 52% and 96% for class 1. The F1 scores are

62% for class 0 and 92% for class 1, with an overall

accuracy of 87%.

The GRU model provides a precision of 82% for

class 0 and 92% for class 1. The recall is 71% for

class 0 and 95% for class 1. The F1 scores are 76%

for class 0 and 94% for class 1, resulting in an overall

accuracy of 90%.

The CNN model reports a precision of 79% for

class 0 and 94% for class 1. The recall values are 79%

for class 0 and 94% for class 1, and the F1 scores are

79% for class 0 and 94% for class 1, with an overall

accuracy of 91%.

In the absence of data augmentation, the CNN-

LSTM model performs well with an accuracy of

92%. Precision is 90% for class 0 and 94% for class

1, demonstrating high accuracy in detecting positive

cases. The recall for class 0 is 85%, while for class 1

it is 92%, indicating effective capture of genuine pos-

itives. The F1 scores are 87% for class 0 and 93% for

class 1, reﬂecting a strong balance between precision

and recall.

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

738

Table 2: Performance metrics without and with augmentation.

Performance Before Augmentation After Augmentation

BERT LSTM GRU CNN CNN-

LSTM

BERT LSTM GRU CNN CNN-

LSTM

Precision 86.5% 87% 83.5% 75% 92% 77% 87% 88% 87.5% 93%

Recall 86.5% 83% 74% 82% 88.5% 81.5% 84.5% 87.5% 87.5%

92.5%

F1 Score 86.5% 80% 77% 77.5% 91.5% 78.5% 86% 88% 88% 92.5%

Accuracy 82% 87% 90% 91% 92% 84% 91% 92% 92% 95%

4.5 Results with Data Augmentation

Data augmentation enhances model performance by

improving precision, recall, and F1 scores, increas-

ing accuracy and generalization. TABLE 2 shows our

comparison results after data augmentation.

Starting with BERT, data augmentation results in a

precision of 61% for class 0 and 93% for class 1. The

recall values are 76% for class 0 and 87% for class 1,

with F1 scores of 67% for class 0 and 90% for class 1.

The overall accuracy of the model improves to 84%.

For LSTM, precision reaches 81% for class 0 and

93% for class 1. Recall values are 74% for class 0

and 95% for class 1, while the F1 scores are 78% for

class 0 and 94% for class 1. The total accuracy of the

model is 91%.

GRU shows a precision of 81% for class 0 and

95% for class 1, with recall values of 80% for class 0

and 95% for class 1. The F1 scores are 81% for class

0 and 95% for class 1, resulting in an overall accuracy

of 92%.

The CNN model achieves a precision of 80% for

class 0 and 95% for class 1. Recall values are 81% for

class 0 and 94% for class 1, with F1 scores of 81% for

class 0 and 95% for class 1. The overall accuracy is

92%.

With data augmentation, the CNN-LSTM model

shows a signiﬁcant performance boost. Precision

rises to 89% for class 0 and 97% for class 1. Recall

improves to 90% for class 0 and 95% for class 1, in-

dicating enhanced detection of true positives. The F1

scores increase to 89% for class 0 and 96% for class

1. Overall accuracy surpasses 95%, reﬂecting a sub-

stantial improvement in model performance due to the

augmentation strategies.

4.6 Performance Visualization

In the visualization section, we evaluate the accuracy

and loss curves of the hybrid LSTM-CNN model, an-

alyze the confusion matrix, and plot the ROC curve to

assess class distinction.

4.6.1 Accuracy Curve

The accuracy curve in Figure 3 illustrates the model’s

performance over 50 training epochs for both the

training and validation datasets. Initially, the train-

ing accuracy increases rapidly, reaching near-perfect

levels around the 10

epoch, where it then plateaus.

The validation accuracy also rises quickly during the

early epochs, stabilizing around 90-92% after the 10

epoch. The gap between training and validation accu-

racy suggests minor overﬁtting, indicating the model

performs well on training data but generalizes less ef-

fectively to new data.

Figure 3: Accuracy Curve.

4.6.2 Loss Curve

The loss curve in Figure 4 represents the model that

initially learns effectively, with both training and vali-

dation losses decreasing rapidly. However, after about

10 epochs, the validation loss plateaus while the train-

ing loss continues to decrease, suggesting that the

model may be overﬁtting to the training data. This di-

vergence implies that the model is becoming too spe-

cialized for the training set, potentially compromising

its ability to generalize well to new, unseen data.

4.6.3 Confusion Matrix

The confusion matrix depicted in Figure 5 summa-

rizes 1,046 instances of Class 0, it correctly predicted

A Hybrid CNN-LSTM Model for Opinion Mining and Classiﬁcation of Course Reviews

739

Figure 4: Loss Curve.

856 and misclassiﬁed 190 as Class 1. For Class 1,

out of 3,796 cases, it accurately predicted 3,711 but

misclassiﬁed 190 as Class 0. While the model shows

a high number of true positives and true negatives,

the presence of false positives (85) and false negatives

(190) indicates areas for improvement in its predictive

capability.

Figure 5: Confusion Matrix.

4.6.4 ROC Curve

The ROC curve shown in Figure 6 model’s ROC

curve is close to the top-left corner, indicating a high

level of classiﬁcation performance. The area under

the curve (AUC) is 0.95, which suggests that the

model has excellent discrimination ability. A higher

AUC value closer to 1.0 implies that the model is very

effective at distinguishing between the positive and

negative classes, with minimal overlap.

4.7 Discussion

The results in the curves and tables underscore the

effectiveness of different models in binary text classi-

ﬁcation tasks, with particular emphasis on the hybrid

CNN-LSTM model. The accuracy and loss curves in-

Figure 6: Roc Curve.

dicate strong learning capabilities, with high training

and consistently robust validation accuracy, suggest-

ing minimal overﬁtting. This is corroborated by the

confusion matrix, which shows the model’s adeptness

at distinguishing between classes, achieving high pre-

cision and recall.

The ROC curve reinforces the model’s perfor-

mance, with a high area under the curve (AUC), in-

dicating strong discrimination between positive and

negative classes. Comparative analysis in TABLE 2

reveals that while BERT and CNN models perform

well, particularly in Class 1, the CNN-LSTM model

provides the best balance of precision, recall, and F1

score, making it the top performer.

The augmented results in TABLE 2 further high-

light the CNN-LSTM model’s superiority, achieving

95% accuracy while maintaining excellent precision

and recall. The improvement with data augmentation

is notable in the LSTM and GRU models, which show

signiﬁcant gains in recall. The CNN-LSTM model re-

mains the most robust, with the highest overall accu-

racy and balanced metrics, underscoring its suitability

for complex classiﬁcation tasks.

5 CONCLUSIONS

In conclusion, this paper presents a comprehensive

study on the automated classiﬁcation of student re-

views in e-learning platforms using a large dataset

from Future Learn. The primary contribution is the

development of a hybrid deep learning architecture

that combines convolutional neural networks (CNN)

and long-short-term memory (LSTM) networks. This

approach achieves remarkable performance, with an

accuracy of 95%, highlighting the robustness of deep

learning in processing and classifying large-scale ed-

ucational data. The dataset’s size and diversity, cou-

pled with the model’s capabilities, underscore its rel-

evance in advancing opinion analysis for online ed-

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

740

ucation and improving e-learning platforms through

sophisticated AI techniques.

Looking ahead, this research lays the groundwork

for enhancing model interpretability and expanding

hybrid architectures to broader educational data min-

ing contexts. Further improvements could include

exploring methods to boost model performance and

scalability. Expanding the dataset may lead to deeper

insights into student feedback, driving enhancements

in e-learning platform effectiveness and user satisfac-

tion.

REFERENCES

Cabada, R. Z., Estrada, M. L. B., and Bustillos, R. O.

(2018). Mining of educational opinions with deep

learning. Journal of Universal Computer Science,

24(11):1604–1626.

Chakravarthy, V., Kameswari, M., Mydeen, H. D., and

Seenivasan, M. (2021). Opinion mining from student

text review for choosing better online courses. In IOP

Conference Series: Materials Science and Engineer-

ing, page 012067. IOP Publishing.

Devlin, J., Chang, M.-W., Lee, K., and Toutanova, K.

(2019). Bert: Pre-training of deep bidirectional trans-

formers for language understanding. In Proceedings

of the 2019 Conference of the North. Association for

Computational Linguistics.

Edalati, M., Imran, A. S., Kastrati, Z., and Daudpota, S. M.

(2022). The potential of machine learning algorithms

for sentiment classiﬁcation of students’ feedback on

mooc. In Intelligent Systems and Applications: Pro-

ceedings of the 2021 Intelligent Systems Conference

(IntelliSys) Volume 3, pages 11–22. Springer.

El-Halees, A. (2011). Mining opinions in user-generated

contents to improve course evaluation. In Software

Engineering and Computer Systems: Second Interna-

tional Conference, ICSECS 2011, Kuantan, Pahang,

Malaysia, June 27-29, 2011, Proceedings, Part II 2,

pages 107–115. Springer.

Guru, C. and Bajnaid, W. (2023). Prediction of cus-

tomer sentiment based on online reviews using ma-

chine learning algorithms. International Journal of

Data Science and Advanced Analytics.

Kastrati, Z., Arifaj, B., Lubishtani, A., Gashi, F., and Nish-

liu, E. (2020a). Aspect-based opinion mining of stu-

dents’ reviews on online courses. In Proceedings of

the 2020 6th International conference on computing

and artiﬁcial intelligence, pages 510–514.

Kastrati, Z., Arifaj, B., Lubishtani, A., Gashi, F., and Nish-

liu, E. (2020b). Aspect-based opinion mining of stu-

dents’ reviews on online courses. In Proceedings of

the 2020 6th International Conference on Comput-

ing and Artiﬁcial Intelligence, ICCAI ’20, pages 510–

514. ACM.

Khder, M. A. (2021). Web scraping or web crawling: State

of art, techniques, approaches and application. Inter-

national Journal of Advances in Soft Computing & Its

Applications, 13(3).

Koufakou, A. (2023). Deep learning for opinion mining and

topic classiﬁcation of course reviews. Education and

Information Technologies, 29(3):2973–2997.

LeCun, Y., Bottou, L., Bengio, Y., and Haffner, P. (1998).

Gradient-based learning applied to document recogni-

tion. Proceedings of the IEEE, 86(11):2278–2324.

MAGNONI, F. and PLUTINO, A. (2018). The move-

me project on the futurelearn platform. some consid-

erations after the ﬁrst pilot. E-Learning, MOOC e

Lingue Straniere: Studi, Ricerche e Sperimentazioni

E-Learning, MOOCs and Foreign Languages: Re-

search, Studies and Experiences, page 103.

Mrhar, K., Benhiba, L., Bourekkache, S., and Abik, M.

(2021). A bayesian cnn-lstm model for sentiment

analysis in massive open online courses moocs. Inter-

national Journal of Emerging Technologies in Learn-

ing (iJET), 16(23):216–232.

Onan, A. (2020). Mining opinions from instructor evalua-

tion reviews: a deep learning approach. Computer Ap-

plications in Engineering Education, 28(1):117–138.

Onan, A. (2021). Sentiment analysis on massive open on-

line course evaluations: a text mining and deep learn-

ing approach. Computer Applications in Engineering

Education, 29(3):572–589.

Pang, B. and Lee, L. (2008). Opinion mining and senti-

ment analysis. Foundations and Trends® in Informa-

tion Retrieval, 2(1–2):1–135.

Shaik, T., Tao, X., Dann, C., Xie, H., Li, Y., and Galligan,

L. (2023). Sentiment analysis and opinion mining on

educational data: A survey. Natural Language Pro-

cessing Journal, 2:100003.

She, D. and Jia, M. (2021). A bigru method for remain-

ing useful life prediction of machinery. Measurement,

167:108277.

Siami-Namini, S., Tavakoli, N., and Namin, A. S. (2019).

The performance of lstm and bilstm in forecasting

time series. In 2019 IEEE International conference

on big data (Big Data), pages 3285–3292. IEEE.

Tripathy, A., Agrawal, A., and Rath, S. K. (2016). Clas-

siﬁcation of sentiment reviews using n-gram machine

learning approach. Expert Systems with Applications,

57:117–126.

Yan, X., Li, G., Li, Q., Chen, J., Chen, W., and Xia, F.

(2021). Sentiment analysis on massive open online

course evaluation. In 2021 International Conference

on Neuromorphic Computing (ICNC), pages 245–249.

IEEE.

A Hybrid CNN-LSTM Model for Opinion Mining and Classiﬁcation of Course Reviews

741