International Journal of

INTELLIGENT SYSTEMS AND APPLICATIONS IN

ENGINEERING
ISSN:2147-67992147-6799 www.ijisae.org Original Research Paper

Indonesian News Classification Using IndoBert

Budi Juarto*1, Yulianto2

Submitted: 12/11/2022 Accepted: 14/02/2023

Abstract: In 2021 there was an increase in the number of people who have internet access, and the number of users increased from 175
million users to 202 million users. News classification in general still uses traditional techniques such as word embedding with TF-IDF
and machine learning. The latest development for technology that can classify NLP news using the BERT model is a state-of-the-art pre-
trained model. However, the pre-train model of BERT is only limited to use in English. So in this study, IndoBERT will be used in
making news recommendations based on the category. This dataset uses an Indonesian news dataset that has 5 categories, including
football, news, business, technology, and automotive. The IndoBERT method will be compared with other pre-train models, such as
XLNET, BERT multilingual, XLMRoberta. Meanwhile, the machine learning method with TF-IDF word embedding was compared
using the XGBoost method, LGB, and random forest. In this study, we see that the classification method using IndoBERT gives the best
results with an accuracy value of 94% and also provides the smallest computation time compared to other methods with a time of one
minute 56 seconds and a validation time of 10 seconds. BERT can give the best results because BERT is a type of pre-trained model that
is trained from various kinds of Indonesian words such as news and several website sources to add to the corpus of vocabulary sources in
the model. In the future research will be carried out to implement the dual IndoBERT model and the Siamese IndoBERT.

Keywords: Natural Language Processing, News Classification, IndoBERT, Multiclass Classification

1. Introduction number is derived from the typical amount of news that is

published in a given year. six days a week. If we
The number of people using the internet in Indonesia has
extrapolate this data further, we find that the daily
consistently shown steady growth from one year to the
readership of print publications is currently at 734
next in their usage. There has been an increase in the
thousand copies, while the daily circulation of digital
number of people using the internet from 175 million users
publications has reached over 2.7 million [3]. The
to 202 million users; this statistic has increased by eleven
categories of news that are held by news organizations are
percent or as many as 27 million users since 2021. The
typically somewhat varied; for example, the Wall Street
number of people using the internet is now 202 million
Journal covers politics, business, and economics in its
users [1]. There are 175 million people who use the
reporting. Because of the vast volume of news, it will be
internet, and of those people, there are 160 million active
challenging to categorize each piece of news according to
users of social media who also use the internet. The present
its respective topic. The incorporation of machine learning
population of Indonesia is 272 million [2], and within that
and deep learning algorithms will make it simpler for
number, there are types of individuals who do not actively
editors of news publications to classify stories following
use the internet, such as toddlers and the elderly.
the appropriate headings.
Therefore, this figure is a significant amount. People that
use the internet in big quantities may find that they can Previous research has been carried out in the instance of
obtain news information via the internet in a way that is Indonesian news using classification using machine
both simple and rapid, increasing the likelihood that they learning XGB utilizing the TF-IDF or a method that counts
will do so. the number of words in a document. [4] In addition, the
XGB approach is utilized to categorize the severity of
According to a source from The Wall Street Journal, the
drought in the Boyolali and Purwerjo regions of Indonesia.
circulation of news online runs quickly and a lot every day,
These regions are located in a part of the country where the
and it has reached an average number of three million and
results of the XGB method are superior to those of random
five hundred thousand readers by the year 2021. This
forests [5]. LightGBM is another machine learning method
that is employed. This study found that methods such as
_______________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________________

1
Computer Science Department, School of Computer Science, Bina
Nusantara University, Jakarta, Indonesia 11480 random forest and also SVM or also called support vector
ORCID ID : 0000-0002-5134-2989 machines have advantages when dealing with large
2
Computer Science Department, School of Computer Science, Bina
Nusantara University, Jakarta, Indonesia 11480 datasets and are suitable for classifying land cover cases in
ORCID ID : 0000-0001-9668-3819 this study [6]
* Corresponding Author Email: [email protected]

International Journal of Intelligent Systems and Applications in Engineering IJISAE, 2023, 11(2), 454–460 | 454
Classification in the field of machine learning in the field were then analyzed using two methods, namely XGBoost
of natural language processing still mostly uses TD-IDF, and Random Forest methods. After evaluating the
which has the disadvantage of being context-independent XGBoost and Random Forest methods, an accuracy value
in making word representations, whereas word of 82% and a kappa value of 65% were obtained, while the
vectorization using BERT is more because it is context random forest method yielded much lower results with an
dependent so that it can provide better word vector accuracy value of 68% and a kappa value of 37%. This
representation [7]. Methods such as TF-IDF are still widely shows that the XGBoost method is better to use than the
used in the field of text classification by combining them Random Forest method because it has a higher accuracy
with several machine learning models where TF-IDF is and kappa value[5].
more focused on the process of converting text into vectors
Shweta D. Mahajan (2021) conducted research on news
so that it can be understood by models in machine learning.
classification using the TF-IDF method for word
When it comes to text categorization, the BERT technique
vectorization and naive-bayes algorithm as a classifier. The
is superior to both TF-IDF and machine learning [8]. In
dataset used is an English dataset with three classes which
addition to this, the BERT method also performs better.
is travel, politics, and criminal classes[11].
The previous study also continues to use BERT as the
classification system for English news [9]. In this Shah et al conducted a study on news classification using
investigation, we will apply IndoBERT to the task of news several machine learning algorithms such as logistic
categorization and evaluate its performance relative to that regression, random forest, and also KNN [12]. Metric
of other pre-trained transformer models. measurements carried out in this study is to use accuracy
as an evaluation metric. In this study it was measured using
2. Related Works an accuracy metric to measure how well the model gave
results in a prediction, the first method using logistic
Machine learning methods have been widely used since the
regression gave results of how much accuracy was 97%,
beginning of research in the field of news classification.
and another method was followed with an accuracy of
After conducting a literature study, it can be seen that
93%, namely using the random forest method. While the
several machine learning methods that are often used are
method that gives the lowest accuracy results is KNN with
SVM, naive Bayes, KNN, random forest, and others.
an accuracy value of only 92%.
Research about comparative study of the SVM, KNN, and
naive-Bayes algorithms in classifying Myanmar language Deep learning methods have increased in popularity in
news. The dataset used has 4 classes, namely politics, recent times due to their performance which can rival and
business, entertainment, and sports with the best beat traditional machine learning methods. Sari et al
performance obtained by the SVM method[10]. conducted a study on the detection of AGNews news using
the LSTM (Long-Short Term Memory) method using
Extreme Gradient Boosting method in classifying
word2vec embeddings[13]. The dataset used in this study
Indonesian news. The classification task carried out is to
is the AGNews dataset which has 4 labels with a total of
distinguish news that is hoax news and valid news. The
496.835 data. The results obtained in this study are the
dataset carried out is a news dataset collected from several
LSTM model with word2vec produces an accuracy of
online news sites. The dataset was collected from 2015 to
95.38% with an average value of precision, recall, and F1-
2020 with the number of studies conducted in the study
score of 95%.
was 500. The distribution of training and test data with a
proportion of 80% training data and 20% testing data. So The traditional method of machine learning has the
that there are 400 news that becomes training data and 100 disadvantage that it requires large labeled data to train it.
that become testing data. In this study, converting each Chenbin Li et al., (2018) used the Bi-LSTM deep learning
word into a vector using the TF-IDF method. The results method combined with CNN to perform text classification
obtained in the study are an accuracy of 92%[4]. tasks [14]. This study uses the word2vec embedding
method to generate input vectors that will be processed by
Research conducted by (Prasetyo et al, 2019) entitled
the model. The developed model was compared with the
"Data Analysis of Landsat 8 OLI Imagery as Drought
TF-IDF, SVM, LSTM and CNN methods and it was
Prediction Index Using Machine Learning in Boyolali and
concluded that the designed model had the best
Purworejo Regencies. The results of this research can be
performance with an F1 of 0.99.
used as a consideration for the government to regulate land
and also water availability to overcome the problems of Berfu Buy¨uk¨ oz e al., (2020) classified socio-political
climate change, drought, and also the problem of supply news data using the ELMo and DistilBERT methods[15].
and demand of water needed by humans to meet their daily As comparisons for the designed method, the Naive Bayes
needs. In this study, the data used were Landsat 8 OLI multinomial and linear support vector are employed. The
satellite imagery in Boyolali and Purworejo districts which research concludes that the DistilBERT technique

International Journal of Intelligent Systems and Applications in Engineering IJISAE, 2023, 11(2), 454–460 | 455
outperforms ELMo in the field of semantic data value
processing. Additionally, the DistilBERT model is 30%
lower in size and 83% quicker in processing speed than
ELMo.

3. Methodology
3.1. Research Data
The dataset used in this investigation is the secondary
dataset retrieved from GitHub by A.Chandra (2020)[16].
News information is gathered from a variety of Indonesian
news outlets. There are 6127 lines of training data and
2627 lines of testing data in the dataset. The dataset has
five categories: news, business, technology, football, and
automobiles. We keep the pre-processing basic by merely
removing stop words and transforming lowercase letters.
Before data is ready to be used, there are some pre-
processing processes before the sentences are used in the
IndoBERT. After pre-processing data the sentence will be
tokenized with IndoBERT tokenizer and then load the
IndoBERT model for the training of the model to evaluate
in data evaluation and data testing. The process in this
study is described in Fig 1.
3.2. Bidirectional Encoder Representations from Fig. 1 Flowchart for IndoBERT Research Process
Transformers (BERT)
BERT is an architectural model developed from
Transformer by taking transformer encoder layers and
stacking them to build a new architecture [17]. The
original BERT consists of 2 types, namely BERT-base
with 12 encoder layers and BERT-large with 24 encoder
layers. Like transformers, BERT is transfer learning and
can be trained with unlabelled data. 2 types of BERT
training methods are carried out simultaneously, namely
mask language prediction and next sentence prediction.
The results of training with large data without labels will
produce a model that has general knowledge about the
relationship between words and sentences. Furthermore,
the model can be fine-tuned with a specific dataset that is
small and has been labeled according to the task to be
completed. BERT Architecture is shown in Figure 2.

Fig. 2 IndoBERT Architecture

BERT uses Transformers which can learn the relationship
between words by using a mechanism called attention.
BERT has an encoder and a decoder in its architecture. The
difference lies in the encoder whose function is to receive
input and turn it into a word vector while the decoder
functions to make predictions from various tasks. To be
able to make words into a vector in the encoder, BERT has
three kinds of embeddings. This is because the BERT
model creates a language model that can be represented

International Journal of Intelligent Systems and Applications in Engineering IJISAE, 2023, 11(2), 454–460 | 456
properly in vector form so that it can provide good models while avoiding their respective limitations. This
predictive model results. Some of the embeddings that are research also proposed a new objective referred to as
carried out in the BERT input include: permutation language modeling, which permits the model
to retain its auto-regression modeling advantage while also
1. Token embeddings: tokens that are at the beginning
capturing bidirectional context. XLNet's neural
of a sentence represented by [CLS] while at the end
architecture is designed to work seamlessly with the
of a sentence represented by tokens by [SEP].
autoregressive objective, including the incorporation of
2. embeddings of segments: This embedding is useful to Transformer-XL and the meticulous design of a two-
be used to distinguish between the two sentences by stream attention mechanism. Yang et al. observe that both
marking each sentence. BERT and XLNet perform partial prediction. In the BERT
3. Positional embeddings: These embeddings provide a case, if all tokens are masked, it is impossible to make
token that is useful for providing positional predictions with any degree of accuracy. In addition, for
information on an embedding in a sentence. both BERT and XL-NET, partial prediction reduces the
difficulty of optimization by predicting only tokens with
sufficient context.
3.5. XLM
Cross-lingual language models (XLMs) are models for
multiple languages that have been trained using two
distinct methodologies. One unsupervised, which uses
solely monolingual data, and one supervised, which uses
Fig. 3 IndoBERT Architecture parallel data and a novel cross-lingual language model
objective[20]. Without the use of a single parallel sentence,
The Transformer, in essence, applies a layer that maps
a cross-lingual language model trained on the XNLI cross-
sequences to sequences, producing an output that is also a
lingual classification benchmark exceeds the previous
series of vectors with input and output tokens
supervised state of the art by an average of 1.3% accuracy.
corresponding 1:1 at the same index. And as we already
know, BERT does not attempt to predict the next word that 3.6. ROBERTA
will be spoken. Two strategies are used during training. The ROBERTA model is intended to optimize the BERT
3.3. IndoBERT model by analyzing the effect of parameters and training
data size[21]. The modification consists of training the
IndoBert is one of the state-of-the-art in conducting text
model with a larger batch but fewer data points. Remove
analysis using Indonesian. The architecture that is built is
the prediction aim for the next sentence. The training uses
using the transformer model on BERT in general which
a longer sequence and dynamically modifies the masking
uses English. As in BERT, this method also uses twelve
pattern applied to the data.
hidden layers where each hidden layer is limited to a
dimension of 786 and also uses 12 attention heads. 3.7. Multilingual BERT
IndoBERT was also built using a large number of Multilingual BERT is a pre-trained model which is
Indonesian vocabulary with a total of more than two continued from BERT which was previously only trained
hundred twenty million words where the sources obtained in English but now is also trained in various languages.
are language sources that use good and correct Indonesian There are 104 languages trained in multilingual BERT
such as from online newspapers and also from the which can provide good annotations between various
Indonesian Web Corpus and other sources. other. languages. Just like the BERT model, this model uses 12
IndoBERT itself is a pre-trained model that has been layers of the transformer model and uses Wikipedia data
developed previously using 2.4 million steps or 180 from various languages, not only from English. BERT has
epochs. The training process takes up to two months. So good results for various languages and Indonesian is one of
that IndoBERT has good performance for several tasks in the languages that has been included in this pre-trained
NLP cases. In testing the IndoLEM dataset, IndoBERT has model. So in this study, we will also try the Multilingual
provided better results than MalayBERT. BERT model to compare its performance to the IndoBERT
3.4. XL-NET model.

Yang et al. (2019) proposed a new model dubbed XLNet

4. Results and Discussion
that takes into account the position of masked words on the
objective of masked language modeling [19]. XLNet This study uses accuracy, precision, and recall evaluation
aimed to optimize auto-regressing and auto-encoding metrics for the evaluation process to be compared to other

International Journal of Intelligent Systems and Applications in Engineering IJISAE, 2023, 11(2), 454–460 | 457
deep learning and machine learning algorithm. These y logy
evaluation metrics are used to predict whether each piece
Actual TP
of news will be grouped in the correct class. These
automotiv Autom
evaluation metrics are some formulas are shown in
e otive
equation 1, equation 2, and equation 3. Some of the terms
used in this equation, such as:
• True Positives (TP_i) – condition if the prediction Based on formulas 1, 2, and 3, we can evaluate the result
result is positive then the actual value is also positive so of IndoBERT compared to machine learning and deep
that the classification is correct. learning model. The result of the evaluation is shown in
Table 2 and Table 3.
• True Negatives (TN_i) – condition if the prediction
result gives a negative label and the actual label is TABLE 2. COMPARISON OF THE EVALUATION RESULTS OF
negative so that the classification is correct. THE MACHINE LEARNING METHOD WITH INDOBERT

• False Positives (FP_i) – condition if the prediction Method Evaluation Metric

results give a negative label even though the actual
Accuracy Precision Recall
value is positive so that the classification gives the
wrong label. XGB 0.89 0.89 0.89

• False Negatives (FN_i) – conditions where the LGBMC 0.92 0.92 0.92
prediction results give positive results but the actual
Random 0.88 0.89 0.88
label is negative so that the classification gives the
Forest
wrong label.
IndoBert 0.95

𝑇𝑃𝑖 +𝑇𝑁𝑖
∑𝑙𝑖=1
𝑇𝑃𝑖 +𝑇𝑁𝑖 +𝐹𝑃𝑖 +𝐹𝑁𝑖
𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = 𝑥100% (1) TABLE 3. COMPARATIVE EVALUATION OF VARIOUS PRE-
𝑙
TRAINED MODELS
∑𝑙𝑖=1 𝑇𝑃𝑖
𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 = ∑𝑙𝑖=1(𝐹𝑃𝑖+ 𝑇𝑃𝑖 )
𝑥100% (2)
Model Training Validation
Accuracy
∑𝑙𝑖=1 𝑇𝑃𝑖
Time Time
𝑅𝑒𝑐𝑎𝑙𝑙 = ∑𝑙𝑖=1(𝑇𝑃𝑖+ 𝐹𝑁𝑖 )
𝑥100% (3)
XLNet 88.4 00:03:10 00:00:20
BERT
In this case, the classification uses 5 news labels so that the Multilingual 91.7 00:01:59 00:00:10
evaluation uses the Confusion Matrix Multiclass. The XLM-
multi-class confusion matrix table used in this study is RoBERTa 91.9 00:02:01 00:00:10
shown in Table 1.
IndoBERT 94.5 00:01:56 00:00:10
TABLE 1 CONFUSION MATRIX FOR NEWS CLASSIFICATION

Predi Predi Predict Predict

cted Predic cted ed ed Table 3 shows the result between IndoBERT and other
Socce ted Busin Techno Autom machine learning method. IndoBERT has higher accuracy
r News ess logy otive than other machine learning methods. In the Table, II
IndoBERT is compared to another deep learning method
TP for multilingual language models, such as XLNet, BERT
Actual Socce Multilingual, IndoBERT, and XLM-RoBERTa. IndoBERT
Soccer r has the lowest training loss, validation loss, and validation
Actual TP time with the highest accuracy than other deep learning
news News methods. The training and validation loss for each epoch is
shown in Figure 4, Figrue 5, Figure 6. and Figure 7. Table
TP 2 and Table 3 show the accuracy summary and IndoBERT
Actual Busin has the highest accuracy with a score of 94.5%.
business ess
Based on the results of this study it can be seen that the
Actual TP IndoBERT method has the best results of all methods
technolog Techno because IndoBERT is trained using more than two hundred

International Journal of Intelligent Systems and Applications in Engineering IJISAE, 2023, 11(2), 454–460 | 458
million words obtained from news sources, Wikipedia, and 5. Conclusions
also from Indonesian Web Corpus website sources so that
We compare IndoBERT with machine learning models
to carry out the classification task IndoBERT has a better
such as XGB, random forest, and light gradient boosting.
pre-trained model than some machine learning
The results show that the accuracy of the IndoBERT model
classification models and other deep learning pre-trained
provides the highest accuracy value, which is 94.5%, while
models.
the machine learning model which has the highest
Fig 4. Training and validation loss XL-Net accuracy in the light gradient boosting model is 92%. Next,
IndoBERT is compared with models such as XLMNet,
XLM Roberta, and multi-language BERT that support
multiple languages such as Indonesian. The results show
that IndoBERT produces the highest accuracy value
compared to the XLMNet, XLM Roberta, and multi-
language BERT models. The IndoBERT model also
produces the smallest training loss and validation loss
values compared to other pre-trained models. The shortest
computation time is in the IndoBERT model compared to
other pre-trained models. IndoBERT provides the best
classification results because IndoBERT's pre-trained
models are trained from Indonesian news.
For future research, the Siamese BERT model can be
developed using dual BERT. The first BERT is used for
news headlines and the second BERT is used for news
content included in the Siamese BERT. Another further
improvement is the IndoBert model can use other BERT
models in Indonesian.

Fig 5. Training and validation loss BERT Multilingual References

[1] A. T. Haryanto, “Riset: Ada 175,4 Juta Pengguna
Internet di Indonesia,” https://inet.detik.com/, 2020.
https://inet.detik.com/cyberlife/d-4907674/riset-ada-
1752-juta-pengguna-internet-di-indonesia.
[2] P. Agustini, “Warganet Meningkat, Indonesia Perlu
Tingkatkan Nilai Budaya di Internet,”
aptika.kominfo.go.id, 2021.
https://aptika.kominfo.go.id/2021/09/warganet-
meningkat-indonesia-perlu-tingkatkan-nilai-budaya-
di-internet/ (accessed Sep. 12, 2021).
Fig 6. Training and validation loss IndoBERT
[3] A. Watson, “Average circulation of the Wall Street
Journal from 2018 to 2020,” www.statista.com, 2020.
https://www.statista.com/statistics/193788/average-
paid-circulation-of-the-wall-street-journal/ (accessed
Jul. 07, 2021).
[4] J. P. Haumahu, S. D. H. Permana, and Y.
Yaddarabullah, “Fake news classification for
Indonesian news using Extreme Gradient Boosting
(XGBoost),” IOP Conf. Ser. Mater. Sci. Eng., vol.
1098, no. 5, p. 052081, 2021, doi: 10.1088/1757-
899x/1098/5/052081.
[5] S. Y. J. Prasetyo, Y. B. Christianto, and K. D.
Fig 7. Training and validation loss XLM ROBERTA Hartomo, “Analisis Data Citra Landsat 8 OLI Sebagai
Indeks Prediksi Kekeringan Menggunakan Machine

International Journal of Intelligent Systems and Applications in Engineering IJISAE, 2023, 11(2), 454–460 | 459
 Learning di Wilayah Kabupaten Boyolali dan [16] A. Chandra, “Indonesian News Dataset,”
Purworejo,” Indones. J. Model. Comput., vol. 2, no. https://github.com/, 2020.
2, pp. 25–36, 2019, [Online]. Available: https://github.com/andreaschandra/indonesian-news.
https://ejournal.uksw.edu/icm/article/view/2954.
[17] J. Devlin, M. W. Chang, K. Lee, and K. Toutanova,
[6] D. A. McCarty, H. W. Kim, and H. K. Lee, “BERT: Pre-training of deep bidirectional
“Evaluation of light gradient boosted machine transformers for language understanding,” NAACL
learning technique in large scale land use and land HLT 2019 - 2019 Conf. North Am. Chapter Assoc.
cover classification,” Environ. - MDPI, vol. 7, no. 10, Comput. Linguist. Hum. Lang. Technol. - Proc.
pp. 1–22, 2020, doi: 10.3390/environments7100084. Conf., vol. 1, no. Mlm, pp. 4171–4186, 2019.
[7] A. K. Singh and M. Shashi, “Vectorization of text [18] [F. Koto, A. Rahimi, J. H. Lau, and T. Baldwin,
documents for identifying unifiable news articles,” “IndoLEM and IndoBERT: A Benchmark Dataset
Int. J. Adv. Comput. Sci. Appl., vol. 10, no. 7, pp. and Pre-trained Language Model for Indonesian
305–310, 2019, doi: 10.14569/ijacsa.2019.0100742. NLP,” pp. 757–770, 2021, doi:
10.18653/v1/2020.coling-main.66.
[8] S. González-Carvajal and E. C. Garrido-Merchán,
“Comparing BERT against traditional machine [19] Z. Yang, Z. Dai, Y. Yang, J. Carbonell, R.
learning text classification,” no. Ml, 2020, [Online]. Salakhutdinov, and Q. V. Le, “XLNet: Generalized
Available: http://arxiv.org/abs/2005.13012. autoregressive pretraining for language
understanding,” Adv. Neural Inf. Process. Syst., vol.
[9] K. S. Nugroho, A. Y. Sukmadewa, and N. Yudistira,
32, no. NeurIPS, pp. 1–11, 2019.
“Large-Scale News Classification using BERT
Language Model: Spark NLP Approach,” 2021, [20] [A. Conneau and G. Lample, “Cross-lingual language
[Online]. Available: http://arxiv.org/abs/2107.06785. model pretraining,” Adv. Neural Inf. Process. Syst.,
vol. 32, 2019.
[10] K. Thandar Nwet, “Machine Learning Algorithms for
Myanmar News Classification,” Int. J. Nat. Lang. [21] J. Libovický, R. Rosa, and A. Fraser, “On the
Comput., vol. 8, no. 4, pp. 17–24, 2019, doi: Language Neutrality of Pre-trained Multilingual
10.5121/ijnlc.2019.8402. Representations,” pp. 1663–1674, 2020, doi:
10.18653/v1/2020.findings-emnlp.150.
[11] S. MAHAJAN, “News Classification Using Machine
Learning,” Int. J. Recent Innov. Trends Comput.
Commun., vol. 9, no. 5, pp. 23–27, 2021, doi:
10.17762/ijritcc.v9i5.5464.
[12] K. Shah, H. Patel, D. Sanghvi, and M. Shah, “A
Comparative Analysis of Logistic Regression,
Random Forest and KNN Models for the Text
Classification,” Augment. Hum. Res., vol. 5, no. 1,
2020, doi: 10.1007/s41133-020-00032-0.
[13] W. K. Sari, D. P. Rini, R. F. Malik, and I. S. B.
Azhar, “Klasifikasi Teks Multilabel pada Artikel
Berita Menggunakan Long Short- Term Memory
dengan Word2Vec,” Resti, vol. 1, no. 10, pp. 276–
285, 2017.
[14] [C. Li, G. Zhan, and Z. Li, “News Text Classification
Based on Improved Bi-LSTM-CNN,” Proc. - 9th Int.
Conf. Inf. Technol. Med. Educ. ITME 2018, pp. 890–
893, 2018, doi: 10.1109/ITME.2018.00199.
[15] B. Büyüköz, A. Hürriyetoğlu, and A. Özgür,
“Analyzing ELMo and DistilBERT on Socio-political
News Classification,” Proc. Work. Autom. Extr.
Socio-political Events from News 2020, no. May, pp.
9–18, 2020, [Online]. Available:
https://www.aclweb.org/anthology/2020.aespen-1.4.

International Journal of Intelligent Systems and Applications in Engineering IJISAE, 2023, 11(2), 454–460 | 460