Comparison of word embedding features using deep learning in sentiment analysis

TELKOMNIKA Telecommunication Computing Electronics and Control
Vol. 23, No. 2, April 2025, pp. 416~425
ISSN: 1693-6930, DOI: 10.12928/TELKOMNIKA.v23i2.26223  416
Journal homepage: http://telkomnika.uad.ac.id
Comparison of word embedding features using deep learning in
sentiment analysis
Jasmir1
, Errissya Rasywir2
, Herti Yani3
, Agus Nugroho2
1
Departement of Computer Engineering, Faculty of Computer Science, Universitas Dinamika Bangsa, Jambi, Indonesia
2
Department of Informatic Engineering, Faculty of Computer Science, Universitas Dinamika Bangsa, Jambi, Indonesia
3
Department of Information System, Faculty of Computer Science, Universitas Dinamika Bangsa, Jambi, Indonesia
Article Info ABSTRACT
Article history:
Received Mar 24, 2024
Revised Dec 4, 2024
Accepted Jan 22, 2025
In this research, we use several deep learning methods with the word
embedding feature to see their effect on increasing the evaluation value of
classification performance from processing sentiment analysis data. The deep
learning methods used are conditional random field (CRF), bidirectional long
short term memory (BLSTM) and convolutional neural network (CNN). Our
test uses social media data from Netflix application user comments. Through
experimentation on different iterations of various deep learning techniques
alongside multiple word embedding characteristics, the BLSTM algorithm
achieved the most notable accuracy rate of 79.5% prior to integrating word
embedding features. On the other hand, the highest accuracy value results
when using the word embedding feature can be seen in the BLSTM algorithm
which uses the word to vector (Word2Vec) feature with a value of 87.1%.
Meanwhile, a very significant change in value increase was obtained from the
FastText feature in the CNN algorithm. After all the evaluation processes were
carried out, the best classification evaluation results were obtained, namely
the BLSTM algorithm with stable values on all word embedding features.
Keywords:
Deep learning
Sentiment analysis
Social media
Text classification
Word embedding
This is an open access article under the CC BY-SA license.
Corresponding Author:
Jasmir
Department of Computer Engineering, Faculty of Computer Science, Universitas Dinamika Bangsa
St. Jendral Sudirman, Tehok, Jambi Selatan, Jambi, Indonesia
Email: ijay_jasmir@yahoo.com
1. INTRODUCTION
In recent decades, technological developments have experienced a rapid surge, especially since the
emergence of the internet and personal computers in the 1980s. These technological advances have caused
major changes in various sectors, including information and communication [1], [2]. The significant increase
in internet technology has expanded the reach of information distribution. One aspect that supports this increase
is social media, where users not only function as recipients of information but also as creators of information.
The increase in the number of internet users in Indonesia is due to the various conveniences offered by social
media and the internet. Through social media, people can access information and communicate very quickly.
The use of data from social media is the latest innovative step that provides an alternative data source
outside of traditional data collection methods [3], [4]. Data collection via social media is considered to provide
efficiency in many ways. This efficiency includes the costs that must be incurred for data acquisition, being
able to obtain data in real time, and producing data that has more detailed information to describe the true
opinion of the community [5]. Activities such as those above that are related to analyzing and responding to
public opinion using data sourced from social media are called sentiment analysis [6], [7].
Sentiment analysis, which is a subset of natural language processing (NLP), uses machine learning
methods to recognize and extract factual information from written text [8]. This analysis involves identifying

TELKOMNIKA Telecommun Comput El Control 
Comparison of word embedding features using deep learning in sentiment analysis (Jasmir)
417
emotional nuances and determining the overall sentiment—whether positive, neutral, or negative—expressed
by the author. Applying sentiment analysis to larger data sets allows for a more comprehensive and in-depth
level of analysis [9].
In NLP, computers do not have an innate understanding of textual language, so they need techniques
to convert words into vectors for easier understanding. The process of representing word vectors remains an
interesting area of research. This representation holds great significance as it profoundly influences the
accuracy and efficacy of the constructed learning models. This word representation technique is included in
the feature engineering section. Feature engineering in textual data has its own challenges due to the
characteristics of unstructured text. The feature engineering strategy for textual data that is popularly used is
known as the word embedding feature [10]–[12].
This word embedding feature is collaborated with several classification methods. There are many
types of classifiers that are commonly used to classify sentiment analysis. The methods that are often used are
machine learning methods [13], [14] and deep learning [15]. In this research, the types of methods used are
deep learning methods, namely conditional random field (CRF) [16], bidirectional long short term memory
(BLSTM) [17], and convolutional neural network (CNN) [18]. CRFs are used to build probabilistic models for
sequential data segmentation and labeling. Because it is conditional, CRF is also used to ensure that inference
is easy to do and also avoids the problem of label bias. BLSTM is used to find out the previous information
process and find out the information process afterward. Meanwhile, CNN is used to see processing capabilities
and evaluate classification performance on text data.
We evaluate the effectiveness of different classification methods by testing their performance using
several types of word representations, namely word to vector (Word2Vec) [19], global vectors for word
representation (GloVe) [20], and FastText [21]. The tests were conducted on a sentiment analysis dataset
consisting of Netflix user comments. Netflix was chosen as the object of study due to its high popularity as a
streaming platform, its large user base, and the variety of content it offers. This makes it a relevant topic for
understanding user preferences for digital entertainment services. Analysis of user sentiment, both positive and
negative, can provide valuable insights into their views on the quality of the service, interface, and content
provided.
Similar studies that have been discussed include by Al-Smadi et al. [22] using several deep learning
methods such as BLSTM-CRF combined with Word2Vec features and producing an F1-score of 66.32%. then
BLSTM CRF combined with FastText features producing an F1-score of 69.98%. Then, Jang et al. [23]
proposed a hybrid model of Bi-LSTM+CNN with Word2Vec, the test results showed that the proposed model
produced more accurate classification results, as well as higher recall and F1 scores, than the multi-layer
perceptron (MLP) model, CNN or individual LSTM and hybrid models. Furthermore, Iftikhar et al. [24]
conducted experiments with several deep learning models combined with several word embedding features such
as CNN+Glove, CNN+Word2Vec, LSTM+Glove, and LSTM+Word2Vec. The results of their research stated
that the results of the combination of deep learning with the word embedding feature produced better
performance. Based on the problems, we conducted research as well as the contribution of this research, namely
to improve the evaluation value of the classification performance of deep learning methods, namely CRF,
BLSTM, and CNN by using word embedding features, namely Word2Vec, GloVe, and FastText as techniques
to improve the evaluation value of deep learning classification performance on machine learning datasets on
social media data from Netflix application user comments.
2. MATERIAL AND METHOD
In order for this research to achieve maximum results, we have compiled a series of important steps
that can produce the right model and not widen the direction in achieving the goal. The steps taken to obtain
results that are in accordance with expectations are compiled in the form of a research framework. The research
framework referred is presented in Figure 1.
2.1. Dataset
The dataset was obtained through a data collection process carried out by crawling. We utilize the
Google Play Scraper Python library. To crawl data, the ID of the application from which data is to be retrieved
is first required. In this case, Netflix has the ID ‘com.netflix.mediaclient’. Furthermore, the selection of the
language in the review is an important step, where this study only considers reviews in Indonesian. After
selecting the language, the selection of reviews is based on the score. In this study, the reviews taken have a
score range of 1 to 5. Furthermore, the order of reviews used is most relevant. The amount of data to be taken
also needs to be determined. The data obtained has several attributes, including: reviewId, username,
userImage, content, score, thumbsUpCount, reviewCreatedVersion, at, replyContent, answeredAt, and
appVersion. However, not all of these attributes are required for this study. Therefore, irrelevant or unused
attributes are removed to simplify the data. There are 4 attributes that will be used, namely username, score,

 ISSN: 1693-6930
TELKOMNIKA Telecommun Comput El Control, Vol. 23, No. 2, April 2025: 416-425
418
date, and content. Figure 2 is a flow diagram of data collection. The data used is a raw dataset that will go
through several pre-processing processes before becoming a dataset that is ready to use.
Figure 1. Research framework
2.2. Preprocessing
After getting the Netflix application user review data, the next step is to carry out the preprocessing
stage before entering the sentiment classification stage. This process is important to ensure that the data used
by sentiment classification models is clean, structured, and ready to use. The preprocessing stages carried out
are data cleaning, case folding, tokenization, stopword removal, stemming, and labeling.
2.3. Word embedding
Every word is depicted as a numerical low-dimensional vector. Word embedding enables the capture
of semantic details from extensive text corpora. These embeddings find application in diverse NLP tasks for
optimal word representation. Notably, several algorithms exist for word embedding, including GloVe,
Word2Vec, and FastText. For this research, we utilize pre-trained models encompassing all three features.
2.3.1. GloVe
GloVe is a co-occurrence matrix-based word representation learning technique that captures semantic
relationships between words in a corpus. GloVe combines global statistics-based approaches (such as
co-occurrence matrices) and local context-based methods (such as Word2Vec) to generate word embeddings
in a vector space, allowing for more effective modeling of linear relationships between words [20], [25].

419
Figure 2. Data collection flow chart
2.3.2. Word2Vec
Word2Vec utilizes the occurrence of words in text to establish connections between them. For instance,
it might associate words like “female” and “male” because they frequently occur in comparable contexts.
Word2Vec operates through two architectural forms: context prediction, which forecasts the surrounding words
based on a given word, and context-based prediction (Bag-of-words), which predicts words given a context.
Essentially, Word2Vec takes a textual corpus as input and generates a word vector as output [19], [26].
2.3.3. FastText
Every word is depicted as a collection of n-gram characters, aiding in capturing the essence of shorter
words and facilitating the embedding’s understanding of word prefixes and suffixes. Each n-gram character is
linked with a vector representation, while words are depicted as the sum of these vector representations.
FastText demonstrates strong performance, enabling rapid model training on extensive datasets and offering
representations for words absent in the training data. In cases where a word is absent during model training, it
can be decomposed into n-grams to acquire its embedding vector [21], [27].
2.4. Deep learning
2.4.1. Conditional random fields
CRFs belong to a class of discriminative models ideally suited for classification tasks wherein the
current classification is impacted by contextual factors or adjacent states [28]. CRF finds application in named
entity recognition [29], part-of-speech tagging, gene prediction, noise reduction, and object detection tasks.
Discriminative models, also known as conditional models, are a subset of models commonly employed in
statistical classification, particularly in supervised machine learning. Discriminative classifiers aim to model
the observed data exclusively, learning classification from provided statistics. Approaches in supervised
learning are typically classified into discriminative models or generative models. Discriminative models, in
contrast to generative models, make fewer assumptions about distributions and place greater reliance on data
quality [30], [31].

 ISSN: 1693-6930
420
2.4.2. Bidirectional long short-term memory
Derived from the recurrent neural network (RNN), BLSTM [32] enhances the RNN architecture by
introducing a “gateway” mechanism to regulate the flow of data. Primarily, the long short-term memory
(LSTM) architecture comprises memory cells along with input gates, output gates, and forget gates. These
elements are structured into a chain-like arrangement composed of RNN modules, which enables the smooth
transfer of memory cells along the chain. Moreover, three separate gates are integrated to oversee and regulate
the inclusion or inhibition of information into the memory cell [33].
2.4.3. Convolutional neural netrwork
The CNN is a form of regulated feed-forward neural network that autonomously learns feature
engineering via the optimization of filters, also known as kernels. Unlike lower layer features, higher layer
features are extracted from a broader context window. CNNs are sometimes called shift invariant or space
invariant artificial neural networks (SIANN) because of their architecture, which involves convolution kernels
or filters with shared weights moving across input features. This movement produces a feature map that is
equivalent to translation. However, despite the terminology, many CNNs are not inherently translation
invariant, mainly because of the downsampling operation applied to the input [34]–[36].
3. RESULTS AND DISCUSSION
This section summarizes the results of the experiments conducted according to the previously planned
research flow. This experiment focuses on analyzing text data from social media using several deep learning
methods combined with word embedding features. Training and testing data are divided with an 80:20 division
scheme. This study tests deep learning methods with various variations of word embedding features. The deep
learning method is applied as an approach to sentiment classification on text data. The types of deep learning
methods used include CRF, BLSTM, and CNN. In addition, the word embedding features used include
Word2Vec, GloVe, and FastText.
Table 1 explains the confusion matrix of CRF with three word embedding features and one without
features. In CRF without features, the results are TP=406, FP=104, FN=91, and TN=299. This means that this
model has a fairly low number of TP compared to the use of Word2Vec, GloVe, and FastText features. In
addition, the FP value is quite high, indicating that the model tends to incorrectly identify negative data as
positive. Then CRF with Word2Vec produces TP=512, FP=98, FN=69, and TN=221. The addition of the
Word2Vec feature significantly increases TP (from 406 to 512), indicating that the model is better able to
recognize positive data correctly. However, FP is still quite high (98), and the number of TN decreases
compared to the model without features. This shows that Word2Vec improves the recognition of positive data
but slightly decreases the ability to recognize negative data. In the CRF with GloVe section, there are results
of TP 0=448, FP=91, FN=81, and TN=280. This indicates that GloVe provides more balanced results than
Word2Vec. TP is lower than Word2Vec, but FP is also lower (91), indicating that the model is better at
minimizing errors in classifying negative data. The number of TNs increases compared to Word2Vec. Then
CRF with FastText which produces TP=473, FP=79, FN=59, and TN=289. It can be seen that FastText
provides the best overall performance. TP and TN increase compared to GloVe, while FP and FN are the lowest
among all models. This shows that FastText is very effective in improving the recognition of positive and
negative data, with the least classification errors.
Table 1. Confusion matrix of CRF
Experiment TP FP FN TN
CRF without feature 406 104 91 299
CRF with Word2Vec 512 98 69 221
CRF with GloVe 448 91 81 280
CRF with FastText 473 79 59 289
Table 2 is a CRF test with 3 word embedding variants and one without word embedding. It can be
seen that the accuracy without using features is lower than the model that uses features. This shows the
importance of the embedding feature. In the CRF with Word2Vec section, there are the best results for Recall,
which means that this model is able to capture more actual positive cases. Meanwhile, CRF with GloVe
produces more stable performance in all metrics, although not the best. Then CRF with FastText gives the best
value in accuracy, precision, and F1-Score. This model is the most optimal in producing correct predictions
and maintaining a balance between precision and recall. These results show that the FastText feature provides
a significant increase in accuracy (8.09%) and recall (8.84%), making it an excellent choice for improving the

421
model’s ability to capture true positives. Precision and F1-Score also increase quite significantly, supporting a
balance between correct positive predictions and the ability to capture positive cases. Focus on the F1-Score
metric, since F1-Score is a metric that combines precision and recall, it is very relevant for cases that require a
balance between the two metrics, especially in classification tasks involving data with an imbalanced class
distribution or cases where the balance between precision and recall is a priority. In these imbalanced datasets
where one class is very dominant, accuracy may appear high because the model can ignore the minority class.
F1-Score addresses this problem by taking the minority class into account. CRF with FastText has the highest
F1-Score, indicating that this feature is optimal for producing a good balance.
Table 2. Comparison of CRF evaluation values with word embedding
Experiment Accuracy Precission Recall F1-Score
CRF without feature 78.33333333 79.60784 81.69014 80.63555
CRF with Word2Vec 81.44444444 83.93443 88.12392 85.97817
CRF with GloVe 80.88888889 83.11688 84.68809 83.89513
CRF with FastText 84.66666667 85.68841 88.90977 87.26937
Table 3 explains the confusion matrix of BLSTM with three word embedding features and one without
features. In BLSTM without features, there are values of TP=481, FP=101, FN=83, and TN=235. This means
that the model without features produces quite good performance, with a TP of 481. However, the FP is quite
high (101), indicating that the model often misclassifies negative data as positive. In addition, the TN value is
lower than the model using features, indicating a weaker ability to recognize negative data. In the BLSTM with
Word2Vec section, there are values of TP=501, FP=97, FN=81, and TN=221. With the addition of the
Word2Vec feature, the number of TP increases to 501, indicating that the model is better able to recognize
positive data correctly than the model without features. However, the FP value is still quite high (97), meaning
that the misclassification of negative data as positive remains quite significant. The decrease in the number of
TN also indicates that negative data recognition is slightly impaired. Next, in BLSTM with GloVe, there are
values of TP=480, FP=99, FN=79, and TN=242. The GloVe feature produces a slightly lower number of TP
than Word2Vec (480 vs. 501), but the FN is also lower (79 vs. 81). In addition, the number of FP is smaller
than Word2Vec (99 vs. 97), and TN increases to 242, indicating that the model is better at recognizing negative
data than Word2Vec. Then BLSTM with FastText produces values of TP=505, FP=66, FN=50, and TN=279.
The FastText model gives the best results among all methods. With the highest TP (505) and the lowest FN
(50), this model is very effective in recognizing positive data. In addition, FP is the lowest (66), and TN is the
highest (279), indicating that this model is also very good at recognizing negative data. This confirms that
FastText improves overall performance.
Table 3. Confusion matrix of BLSTM
BLSTM without feature 481 101 83 235
BLSTM with Word2Vec 501 97 81 221
BLSTM with GloVe 480 99 79 242
BLSTM with FastText 505 66 50 279
Table 4 is a summary table of the experimental results of the BLSTM method with three variants of
word embedding features and one without features. In the BLSTM without feature section; this model serves
as a baseline, and its performance is relatively good without additional features, but it can still be further
improved by adding word embedding features. In the BLSTM with Word2Vec section, the model is slightly
better than the baseline, with small improvements in accuracy, precision, recall, and F1-Score. The use of
Word2Vec as an embedding feature improves the model’s understanding of word relationships. In the BLSTM
with GloVe section, the results are very similar to Word2Vec, but slightly lower in precision. This model shows
better performance in terms of Recall, but not as good as the model with Word2Vec. Then in the BLSTM with
FastText section: This is the best model, with significant improvements in all metrics. FastText provides clear
improvements in precision, recall, and F1-Score, making it a very effective model in this classification.
The use of embedding features such as Word2Vec, GloVe, and FastText affects the improvement of
model performance compared to the baseline model without features. The model with FastText shows the
greatest improvement, especially in recall and F1-Score. BLSTM with FastText has the highest F1-Score
(89.69%), indicating that this model is the best choice especially in the balance between accurate prediction
and the model’s ability to capture positive cases. The use of embedding features such as FastText can
significantly improve performance compared to not using features or using other features such as Word2Vec
and GloVe.

 ISSN: 1693-6930
422
Table 4. Comparison of BLSTM evaluation values with word embedding
BLSTM without feature 79.55555556 82.64605 85.28369 83.94415
BLSTM with Word2Vec 80.22222222 83.77926 86.08247 84.91525
BLSTM with GloVe 80.22222222 82.90155 85.86762 84.35852
BLSTM with FastText 87.11111111 88.44133 90.99099 89.69805
Table 5 explains the confusion matrix of CNN with three word embedding features and one without
features. In CNN without features, there are values of TP=404, FP=128, FN=119, and TN=249. Without word
embedding features, CNN produces the lowest performance. The number of TP is the lowest (404), while FP
and FN are the highest (128 and 119). This shows that the model has many errors in recognizing both positive
and negative data. In addition, the TN value is also quite low compared to the model with features. Next, the
CNN with Word2Vec section has TP=512, FP=98, FN=70, and TN=220. The addition of Word2Vec
significantly increases the number of TP to 512, indicating that the model is better able to recognize positive
data than without features. However, the FP (98) and FN (70) values are still quite high, which means there is
room for improvement in recognizing negative data. The decrease in the number of TNs compared to without
features also shows that the model is slightly less effective in recognizing negative data. In the CNN with
GloVe section, there are values of TP=446, FP=88, FN=87, and TN=279. This means that GloVe provides
more balanced results than Word2Vec. FP decreases to 88, while TN increases significantly to 279, indicating
a better ability to recognize negative data. However, TP is lower than Word2Vec (446 vs. 512), and FN is
slightly higher than Word2Vec. Furthermore, CNN with FastText produces values of TP=463, FP=76, FN=74,
and TN=287. FastText produces the best results among all methods. With high TP (463) and low FN (74), the
model is very effective in recognizing positive data. In addition, FP is the lowest (76), and TN is the highest
(287), indicating that this model is also very good at recognizing negative data.
Table 5. Confusion matrix of CNN
CNN without feature 404 128 119 249
CNN with Word2Vec 512 98 70 220
CNN with GloVe 446 88 87 279
CNN with FastText 463 76 74 287
Table 6 is a summary table of the results of CNN experiments with three variants of word embedding
features and one without features. This model shows the best improvement compared to the baseline model.
FastText provides a very good balance between precision, recall, and F1-Score, with excellent results in all
metrics. The use of embedding features such as Word2Vec, GloVe, and FastText significantly improves model
performance compared to the baseline model that does not use additional features. The CNN with FastText
model has the highest F1-Score (86.06%), which shows an optimal balance between prediction accuracy and
the ability to capture positive cases, indicating that this model is very effective in balancing both aspects. The
use of F1-Score in this case is because we want to maintain a balance between accuracy and precision and the
model’s ability to find all positive classes. The CNN with FastText model is the best choice for this model,
with significant improvements in all metrics, especially in recall and precision. Thus, FastText provides better
results than other embedding features such as Word2Vec and GloVe in optimizing text classification
performance.
Table 6. Comparison of CNN evaluation values with word embedding
CNN without feature 72.55555556 75.93985 77.24665 76.58768
CNN with Word2Vec 81.33333333 83.93443 87.97251 85.90604
CNN with GloVe 80.55555556 83.5206 83.6773 83.59888
CNN with FastText 83.33333333 85.89981 86.21974 86.05948
This study examines the impact of performance improvements, computationally BLSTM is very
efficient, this is because the BLSTM process occurs sequentially and regularly, making it suitable for
processing long texts and large datasets. With the word embedding feature, BLSTM can capture more
interactions between features that may be ignored by CRF and CNN. While previous studies have investigated
the impact of other features of the same method. the study did not explicitly discuss their effect on
computational performance.

423
Based on the results of the three experiments, the BLSTM algorithm achieved the highest accuracy of
79.5%, while the CNN algorithm recorded the lowest accuracy of 72.5% before the word embedding feature
was applied. After combining word embedding, BLSTM with the Word2Vec feature achieved the highest
accuracy of 87.1%, while the lowest accuracy post-embedding was also seen in BLSTM using the GloVe and
FastText features. By reviewing all classification evaluation metrics—accuracy, precision, recall, and F1
score—BLSTM emerged as the best performing algorithm, consistently producing stable results across all
embeddings.
However, all tests still allow some false positives and false negatives, indicating potential areas for
further research, such as minimizing these errors. Additional accuracy improvements can be achieved by tuning
hyperparameters. An important observation is that, before embedding, CNN has the lowest performance, but
after applying embedding, especially Word2Vec with BLSTM, the performance improves significantly. This
may be due to the characteristics of CNN which are not well suited for text data, while BLSTM, which reads
sequences bidirectionally, shows a high ability to process text in detail, resulting in superior performance.
4. CONCLUSION
Our study has highlighted the efficacy of pre-trained word embedding models in sentiment analysis.
Through a series of experiments, we have demonstrated the ability of these models to achieve high levels of
accuracy across diverse textual datasets. In our evaluation, various deep learning methods with different word
embedding features were tested with CRF, BLSTM, and CNN algorithms. The use of word embedding features
such as FastText, Word2Vec, and GloVe consistently improved the performance of various text classification
models on CRF, BLSTM, and CNN compared to models without features. FastText was identified as the best
feature based on the table results as it produced the most balanced classification with minimal error. FastText
also produced highly accurate classification on both positive and negative data. Word2Vec excelled in
recognizing positive data but tended to be less accurate on negative data. For limited computational resources,
GloVe can be chosen as it provides balanced results with lower error compared to Word2Vec. GloVe offers a
good balance with lighter computational requirements, suitable for reducing errors on negative data. The choice
of word embedding features used can be tailored to the specific needs of the model and the classification
objectives.
ACKNOWLEDGEMENTS
We would like to thank Yayasan Dinamika Bangsa Jambi for the moral and financial support in
completing this research, and would like to thank the research and community service institution, Universitas
Dinamika Bangsa Jambi for its facilities and annual work programs.
REFERENCES
[1] A. L. Guzman and S. C. Lewis, “Artificial intelligence and communication: A Human–Machine Communication research agenda,”
New Media & Society, vol. 22, no. 1, pp. 70–86, 2020, doi: 10.1177/1461444819858691.
[2] B. Jimada-Ojuolape and J. Teh, “Impact of the Integration of Information and Communication Technology on Power System
Reliability: A Review,” IEEE Access, vol. 8, pp. 24600–24615, 2020, doi: 10.1109/ACCESS.2020.2970598.
[3] R. Lozano-Blasco, M. Mira-Aladrén, and M. Gil-Lamata, “Social media influence on young people and children: Analysis on
Instagram, Twitter and YouTube,” Comunicar, vol. 30, no. 74, pp. 117–128, 2023, doi: 10.3916/C74-2023-10.
[4] B. T. K., C. S. R. Annavarapu, and A. Bablani, “Machine learning algorithms for social media analysis: A survey,” Computer
Science Review, vol. 40, p. 100395, 2021, doi: 10.1016/j.cosrev.2021.100395.
[5] S. M. Fernández-Miguélez, M. Díaz-Puche, J. A. Campos-Soria, and F. Galán-Valdivieso, “The impact of social media on restaurant
corporations’ financial performance,” Sustainability, vol. 12, no. 4, pp. 1–14, 2020, doi: 10.3390/su12041646.
[6] H. R. Alhakiem and E. B. Setiawan, “Aspect-Bas1ed Sentiment Analysis on Twitter Using Logistic Regression with FastText Feature
Expansion,” Jurnal RESTI (Rekayasa Sistem dan Teknologi Informasi), vol. 6, no. 5, pp. 840–846, 2022, doi: 10.29207/resti.v6i5.4429.
[7] M. Birjali, M. Kasri, and A. Beni-Hssane, “A comprehensive survey on sentiment analysis: Approaches, challenges and trends,”
Knowledge-Based Systems, vol. 226, p. 107134, 2021, doi: 10.1016/j.knosys.2021.107134.
[8] A. Palanivinayagam, C. Z. El-Bayeh, and R. Damaševičius, “Twenty Years of Machine-Learning-Based Text Classification: A
Systematic Review,” Algorithms, vol. 16, no. 5, pp. 1–28, 2023, doi: 10.3390/a16050236.
[9] G. Xu, Y. Meng, X. Qiu, Z. Yu, and X. Wu, “Sentiment analysis of comment texts based on BiLSTM,” IEEE Access, vol. 7, pp.
51522–51532, 2019, doi: 10.1109/ACCESS.2019.2909919.
[10] J. Jasmir, W. Riyadi, S. R. Agustini, Y. Arvita, D. Meisak, and L. Aryani, “Bidirectional Long Short-Term Memory and Word
Embedding Feature for,” Jurnal RESTI (Rekayasa Sistem dan Teknologi Informasi), vol. 6, no. 4, pp. 505–510, 2022, doi:
10.29207/resti.v6i4.4005.
[11] S. Ruder, I. Vulić, and A. Søgaard, “A Survey of Cross-lingual Word Embedding Models,” Journal of Artificial Intelligence
Research, vol. 65, pp. 569–630, Aug. 2019, doi: 10.1613/jair.1.11640.
[12] Z. Zhuang, Z. Liang, Y. Rao, H. Xie, and F. L. Wang, “Out-of-vocabulary word embedding learning based on reading
comprehension mechanism,” Natural Language Processing Journal, vol. 5, no. August, p. 100038, 2023, doi:
10.1016/j.nlp.2023.100038.
[13] S. Rapacz, P. Chołda, and M. Natkaniec, “A method for fast selection of machine-learning classifiers for spam filtering,”

 ISSN: 1693-6930
424
Electronics, vol. 10, no. 17, 2021, doi: 10.3390/electronics10172083.
[14] F. N. N. H. R. Passarella, S. Nurmaini, M. N. Rachmatullah, and H. Veny, “Development of a machine learning model for predicting
abnormalities of commercial airplanes,” Data Science and Management, p. 100137, 2023, doi: 10.1016/j.jsamd.2023.100613.
[15] A. Tavanaei, M. Ghodrati, S. R. Kheradpisheh, T. Masquelier, and A. Maida, “Deep learning in spiking neural networks,” Neural
Networks, vol. 111, pp. 47–63, 2019, doi: 10.1016/j.neunet.2018.12.002.
[16] J. Jasmir, S. Nurmaini, R. F. Malik, and B. Tutuko, “Bigram feature extraction and conditional random fields model to improve text
classification clinical trial document,” TELKOMNIKA (Telecommunication, Computing, Electronics and Control), vol. 19, no. 3,
pp. 886–892, 2021, doi: 10.12928/TELKOMNIKA.v19i3.18357.
[17] G. Liu and J. Guo, “Bidirectional LSTM with attention mechanism and convolutional layer for text classificatio,” Neurocomputing,
2019, doi: 10.1016/j.neucom.2019.01.078.
[18] M. Akbar, S. Nurmaini, and R. U. Partan, “The deep convolutional networks for the classification of multi-class arrhythmia,”
Bulletin of Electrical Engineering and Informatics, vol. 13, no. 2, pp. 1325–1333, 2024, doi: 10.11591/eei.v13i2.6102.
[19] R. Rahmanda and E. B. Setiawan, “Word2Vec on Sentiment Analysis with Synthetic Minority Oversampling Technique and
Boosting Algorithm,” Jurnal RESTI (Rekayasa Sistem dan Teknologi Informasi), vol. 6, no. 4, pp. 599–605, 2022, doi:
10.29207/resti.v6i4.4186.
[20] A. George, H. B. B. Ganesh, M. A. Kumar, and K. P. Soman, Significance of global vectors representation in protein sequences
analysis, Springer International Publishing, vol. 31, 2019, doi: 10.1007/978-3-030-04061-1_27.
[21] I. N. Khasanah, “Sentiment Classification Using fastText Embedding and Deep Learning Model,” Procedia CIRP, vol. 189, pp.
343–350, 2021, doi: 10.1016/j.procs.2021.05.103.
[22] M. Al-Smadi, B. Talafha, M. Al-Ayyoub, and Y. Jararweh, “Using long short-term memory deep neural networks for aspect-based
sentiment analysis of Arabic reviews,” International Journal of Machine Learning and Cybernetics, vol. 10, no. 8, pp. 2163–2175,
2019, doi: 10.1007/s13042-018-0799-4.
[23] B. Jang, M. Kim, G. Harerimana, S. U. Kang, and J. W. Kim, “Bi-LSTM model to increase accuracy in text classification:
Combining word2vec CNN and attention mechanism,” Applied Sciences, vol. 10, no. 17, p. 5841, 2020, doi: 10.3390/app10175841.
[24] S. Iftikhar, B. Alluhaybi, M. Suliman, A. Saeed, and K. Fatima, “Amazon products reviews classification based on machine learning,
deep learning methods and BERT,” TELKOMNIKA (Telecommunication Computing Electronics and Control), vol. 21, no. 5, pp.
1084–1101, 2023, doi: 10.12928/TELKOMNIKA.v21i5.24046.
[25] N. Badri, F. Kboubi, and A. H. Chaibi, “Combining FastText and Glove Word Embedding for Offensive and Hate speech Text
Detection,” Procedia Computer Science, vol. 207, pp. 769–778, 2022, doi: 10.1016/j.procs.2022.09.132.
[26] D. Jatnika, M. A. Bijaksana, and A. A. Suryani, “Word2vec model analysis for semantic similarities in English words,” Procedia
Computer Science, vol. 157, pp. 160–167, 2019, doi: 10.1016/j.procs.2019.08.153.
[27] M. A. Raihan and E. B. Setiawan, “Aspect Based Sentiment Analysis with FastText Feature Expansion and Support Vector Machine
Method on Twitter,” Jurnal RESTI (Rekayasa Sistem dan Teknologi Informasi), vol. 6, no. 4, pp. 591–598, 2022, doi:
10.29207/resti.v6i4.4187.
[28] Q. Zhang, Y. Cao, and H. Yu, “Parsing citations in biomedical articles using conditional random fields,” Computers in biology and
medicine, vol. 41, no. 4, pp. 190–194, 2011, doi: 10.1016/j.compbiomed.2011.02.005.
[29] W. Lee, K. Kim, E. Y. Lee, and J. Choi, “Conditional random fields for clinical named entity recognition: A comparative study
using Korean clinical texts,” Computers in Biology and Medicine, vol. 10, pp. 7–14, 2018, doi: 10.1016/j.compbiomed.2018.07.019.
[30] P. Corcoran, P. Mooney, and M. Bertolotto, “Linear street extraction using a Conditional Random Field model,” Spatial
Statisticsvol, vol. 14, pp. 532–545, 2015, doi: 10.1016/j.spasta.2015.10.003.
[31] C. Jiang, M. Maddela, W. Lan, Y. Zhong, and W. Xu, “Neural CRF model for sentence alignment in text simplification,”
Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, 2020, pp. 7943–7960, doi:
10.18653/v1/2020.acl-main.709.
[32] K. S. Tai, R. Socher, and C. D. Manning, “Improved Semantic Representations From Tree-Structured Long Short-Term Memory
Networks,” arXiv preprint, 2015, doi: 10.48550/arXiv.1503.00075.
[33] Z. Dai, X. Wang, P. Ni, Y. Li, G. Li, and X. Bai, “Named Entity Recognition Using BERT BiLSTM CRF for Chinese Electronic
Health Records,” 2019 12th International Congress on Image and Signal Processing, BioMedical Engineering and Informatics
(CISP-BMEI), Suzhou, China, 2019, pp. 1–5, doi: 10.1109/CISP-BMEI48845.2019.8965823.
[34] D. T. Putra and E. B. Setiawan, “Sentiment Analysis on Social Media with Glove Using Combination CNN and RoBERTa,” Jurnal
RESTI (Rekayasa Sistem dan Teknologi Informasi), vol. 7, no. 3, pp. 457–563, 2023, doi: 10.29207/resti.v7i3.4892.
[35] J. Zhang, F. Liu, W. Xu, and H. Yu, “Feature fusion text classification model combining CNN and BiGRU with multi-attention
mechanism,” Future Internet, vol. 11, no. 11, 2019, doi: 10.3390/fi11110237.
[36] J. Yao, C. Wang, C. Hu, and X. Huang, “Chinese Spam Detection Using a Hybrid BiGRU-CNN Network with Joint Textual and
Phonetic Embedding,” Electronics, vol. 11, no. 15, 2022, doi: 10.3390/electronics11152418.
BIOGRAPHIES OF AUTHORS
Jasmir is senior lecture at Universitas Dinamika Bangsa Jambi, Indonesia. He
received his Bachelor in Computer Engineering in 1995 and Master degree in Information
Technology in 2006 from Universitas Putra Indonesia YPTK Padang, Indonesia. He receives
a Doctor in Informatics Engineering at Universitas Sriwijaya Palembang, Indonesia in 2022.
His research interest is data mining, machine learning and deep learning for natural language
processing, and its application. He can be contacted at email: ijay_jasmir@yahoo.com.

425
Errissya Rasywir received the Bachelor degree (S.Kom) in Computer Science
from the Sriwijaya University. She received the Master degree (M.T) in Informatics Master
STEI from the Institut Teknologi Bandung (ITB). She is a lecture of computer science in the
Informatics Engineering, Dinamika Bangsa University (UNAMA). She is currently studying
for a Doctorate in Computer Science at Sriwijaya University. In addition, she is serving as
head of the research group (LPPM) on UNAMA. Her research interests are in data mining,
artificial intelligent (AI), natural languange proccessing (NLP), machine learning, and deep
learning. She can be contacted at email: errissya.rasywir@gmail.com.
Herti Yani is a lecture at Universitas Dinamika Bangsa Jambi, Indonesia. She
received his Bachelor in Information System in Universitas Dinamika Bangsa Jambi in 2009
and Master degree in Magister System Information in Universitas Dinamika Bangsa Jambi,
Indonesia in 2011. She is currently studying for a Doctorate in Computer Science at Satya
Wacana Christian University. Her research interest are in database, artificial intelligence, and
machine learning. She can be contacted at email: adeherti@unama.ac.id.
Agus Nugroho is lecture at Universitas Dinamika Bangsa Jambi, Indonesia. He
received his Bachelor in Informatics Engineering in Universitas Dinamika Bangsa Jambi in
2011 and Master degree in Magister of Informatics Engineering in STMIK AMIKOM
Yogyakarta, Indonesia in 2013. His research interest are in multimedia, artificial intelligence,
and machine learning. She can be contacted at email: agusnugroho0888@gmail.com.

Comparison of word embedding features using deep learning in sentiment analysis

More Related Content

Similar to Comparison of word embedding features using deep learning in sentiment analysis (20)

More from TELKOMNIKA JOURNAL (20)

Recently uploaded (20)

Comparison of word embedding features using deep learning in sentiment analysis