The IJICS (International Journal of Informatics and Computer Science)
Vol 9 No 2, July 2025, Page 98-108
ISSN 2548-8384 (online), ISSN 2548-8449 (print)
Available Online at [Link]
DOI 10.30865/ijics.v9i2 .8917
Implementation of Deep Learning Method Using BERT Model in Career
Choice Analysis of Gen Z
Silvia Ramadhani*,Abdul Halim Hasugian
Computer Science, State Islamic University of North Sumatra, Medan, Indonesia
Email: 1*slvrmdhni11@[Link], 2abdulhalimhasugian@[Link]
(* : slvrmdhni11@[Link])
Submitted: 22/07/2025; Accepted: 30/07/2025; Published: 31/07/2025
Abstract−The development of digital technology has significantly influenced how individuals, particularly Generation Z (born
between 1997 and 2012), make career decisions. Faced with an abundance of digital information, many individuals in this cohort
experience difficulties in selecting career paths that align with their interests, abilities, and labor market demands. This study analyzes
the career preferences of Generation Z using a deep learning approach through the Bidirectional Encoder Representations from
Transformers (BERT) model, specifically the IndoBERT variant, which is pre-trained on Indonesian-language data. The research data
were collected from textual responses to Google Form questionnaires, focusing on four digital career paths: Software Engineer, Content
Creator, Digital Marketing, and Entrepreneur. From 601 data samples, sentiment analysis revealed that 57.85% of the responses were
positive, while 42.15% were negative. Classification results indicated that Content Creator was the most preferred career, followed by
Entrepreneur, Digital Marketer, and Software Engineer. Model evaluation showed a test accuracy of 51.24%, with better performance
in categories that had larger data volumes. These findings demonstrate that IndoBERT is effective in capturing opinions and career
tendencies from unstructured text and provides a scientific basis for educational institutions, industries, and policymakers to design
more relevant career development strategies in the digital era.
Keywords: Generation Z; career choice; BERT; sentiment analysis; deep learning
1. INTRODUCTION
The rapid development of digital technology has influenced various aspects of life, including how individuals
especially Generation Z determine their career paths[1]. Born between 1997 and 2012, this generation has grown up in
an environment saturated with the internet, social media, and mobile technology[2]. Despite having broad access to
information, Gen Z often faces difficulties in choosing careers that align with their interests, abilities, and labor market
demands. In today’s digital era, they tend to be drawn to professions related to technology and creativity, such as Software
Engineer, Content Creator, Digital Marketing, and Entrepreneur. These fields are considered attractive due to their
flexibility, high income potential, and strong connection to the rapidly evolving digital ecosystem.
There are four main factors that influence the career decisions of Generation Z: work conditions and environment,
job prospects and duration, intrinsic values and contribution to work, and work schedule flexibility. However, previous
research remains limited to conventional quantitative survey approaches[3]. Generation Z is a productive age group that
is beginning to enter the workforce and is expected to become the main driving force of future labor markets. They tend
to prefer careers that offer flexibility, creativity, and strong connections to technology and social media. However, the
abundance of digital information often leads to confusion in choosing careers that align with their interests and market
demands[4]. Therefore, a systematic and data-driven analysis is needed to understand Gen Z’s career preferences. The
results of such analysis can serve as a foundation for educational institutions, companies, and policymakers to design
more targeted career development strategies and to bridge the gap between current and future workforce needs.
This study aims to enhance that approach by applying the BERT deep learning model, which is capable of
analyzing unstructured textual data in a deep and contextual manner, particularly from social media. This method is
expected to provide a more comprehensive understanding of Generation Z’s career preferences within the digital
ecosystem.
Sentiment analysis is a method used to identify and categorize opinions or emotions expressed in digital text,
aiming to determine the author’s attitude toward a particular topic whether it is positive, negative, or neutral. This method
is commonly applied to product reviews, news articles, or social media content to understand public opinion trends and
consumer perceptions, which are then used as a basis for decision-making[5]. Sentiment analysis, also known as opinion
mining, is a field of study in computing that aims to identify and represent opinions, sentiments, evaluations, emotions,
subjectivity, attitudes, and viewpoints contained within a text. This method helps in understanding how an individual or
a group responds to a particular topic through the language they use[6].
Machine learning is a branch of artificial intelligence that can be defined as a method that enables systems to learn
from experience, adapt to new data, perform actions automatically, and make accurate and relevant decisions. This
technology plays a role in supporting decision-making processes, where systems can learn from available data, recognize
patterns, and automatically generate decisions without direct human intervention[7]. Machine learning is a branch of
computer science that develops algorithms capable of recognizing patterns in data and learning automatically without
being explicitly programmed. This technology allows systems to become more adaptive and efficient[8].
Deep learning is a branch of machine learning that relies on artificial neural networks with multiple layers (deep
neural networks) to process and understand complex data. Each layer plays a role in extracting and interpreting
information from the data at different levels[9]. Deep learning utilizes complex models for various applications such as
Copyright © 2025, Silvia Ramadhani. This is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. Page 98
The IJICS (International Journal of Informatics and Computer Science)
Vol 9 No 2, July 2025, Page 98-108
ISSN 2548-8384 (online), ISSN 2548-8449 (print)
Available Online at [Link]
DOI 10.30865/ijics.v9i2 .8917
speech recognition, natural language processing, and image recognition, supported by large amounts of data and high
computational power. Its main advantage is the ability to automatically learn from unstructured or unlabeled data, making
it highly effective for solving complex tasks that are difficult to address using traditional learning methods. As a result,
deep learning has become a crucial technology in fields such as artificial intelligence, big data analytics, and the
development of advanced technologies[10].
BERT (Bidirectional Encoder Representations from Transformers) is an encoder model based on the Transformer
architecture that uses attention mechanisms to understand the contextual relationships between words, even those that are
far apart, without relying on recurrent networks. BERT generates contextualized representations for each word in a
sentence and has been shown to significantly improve model performance on various complex and sequential NLP
tasks[11]. Natural Language Processing (NLP) is a branch of artificial intelligence in which computers are designed to
communicate with humans using natural language, such as Indonesian[12].
This study aims to analyze the sentiments and perceptions of Generation Z toward four major digital career paths
Software Engineer, Content Creator, Digital Marketing, and Entrepreneur by utilizing the BERT algorithm on
questionnaire data collected via Google Forms. Through this approach, the research seeks to identify career preferences
among Generation Z and the factors influencing their choices. The results of this analysis can serve as a valuable reference
for educational institutions, companies, and career counselors in designing more targeted and relevant career development
programs suited to the needs of the younger generation in the digital era.
Previous research utilized the BERT model to analyze sentiment in hotel reviews, primarily focusing on classifying
customer opinions into categories such as positive and negative. In contrast, this study applies BERT to analyze career
choices among Generation Z, involving more complex and diverse data related to individual perceptions, motivations,
and preferences regarding careers. The analytical objective also differs; rather than simply performing classification, it
aims to understand the context and patterns behind career decision-making. This highlights a clear research gap, as there
is limited exploration of BERT's application in the field of career development, particularly among Indonesian youth. By
addressing this gap, this study contributes both methodologically and practically: methodologically by extending the use
of NLP-based sentiment analysis into sociological and psychological domains, and practically by providing insights that
can inform educational institutions and policymakers in shaping future career guidance strategies. Thus, this study
expands the application of BERT beyond commercial use to include generational studies and decision-making processes.
Previous research utilized BERT Uncased and IndoBERT for emotion classification in Indonesian text using a
dataset of 2,515 entries with nine emotion labels. The results showed that BERT Uncased achieved 90% accuracy, while
IndoBERT reached 81%, with the best performance observed at the 10th epoch. IndoBERT proved effective in identifying
emotions, achieving up to 97.6% accuracy on opinion sentences[13]. Another study titled "Hoax News Detection through
Website Comparison Using a Deep Learning Approach with the BERT Algorithm" compared the performance of three
algorithms BERT, SVM, and Random Forest in classifying hoax news. The results showed that BERT significantly
outperformed the others, achieving both an accuracy and ROC-AUC score of 0.99. This advantage is attributed to BERT’s
ability to deeply process contextual information through its multi-layered architecture. Additionally, the text analysis
revealed that hoax news is often accompanied by manipulated visuals, such as images or videos that do not match the
actual content[14].
This research is important because Generation Z is poised to dominate the workforce, and their career preferences
will directly shape future labor market trends. Moreover, by leveraging advanced NLP techniques, this study bridges the
methodological gap between computational approaches and career development research in Indonesia.
2. RESEARCH METHODOLOGY
2.1 Research Framework
Based on this study, there are several research stages. The research was conducted using data from social media.
The image below illustrates the systematic stages of the research process. It begins with the Planning phase, followed by
Data Collection (gathering data from social media). Next, the Method Analysis stage is carried out to determine the
appropriate approach, which is then followed by the Application of Method (implementation of the IndoBERT method).
The Testing Method phase is conducted to evaluate the model using the processed data. The outcomes are presented in
the Results stage, and finally, an Evaluation is performed to assess the overall performance of the model. This
visualization provides a clear and structured overview of the research workflow implemented in this study.
Copyright © 2025, Silvia Ramadhani. This is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. Page 99
The IJICS (International Journal of Informatics and Computer Science)
Vol 9 No 2, July 2025, Page 98-108
ISSN 2548-8384 (online), ISSN 2548-8449 (print)
Available Online at [Link]
DOI 10.30865/ijics.v9i2 .8917
Figure 1. Research Framework
2.1.1 Planning
This study begins with data collection through open-ended questionnaires distributed online to Generation Z to
explore their career preferences and the factors influencing their career choices in the digital era. The respondents' textual
answers go through preprocessing stages such as data cleaning, normalization, stopword removal, and stemming. The
processed data is then analyzed using the IndoBERT model to identify patterns or trends in Gen Z’s career choices more
deeply with the help of NLP technology[15].
2.1.2 Data Collection
Data collection in this study was carried out through the design and online distribution of open-ended
questionnaires to Generation Z, aiming to explore their career choices and the influence of the digital environment. The
free-text responses were gathered in digital format and initially reviewed to ensure completeness and relevance before
being analyzed using the BERT model. This approach aims to directly capture Gen Z’s career preferences in the digital
era.
2.1.3 Method Analysis
The method analysis section focuses on processing text data from questionnaires using the BERT deep learning
model, specifically IndoBERT, which is tailored for the Indonesian language[16]. The data is first processed through
several preprocessing stages, including data cleaning, normalization, stopword removal, and stemming, followed by
tokenization using the IndoBERT tokenizer. After preprocessing, the model is trained to identify patterns in Gen Z's
career preferences and evaluated using performance metrics such as accuracy, precision, recall, and F1-score. The ultimate
goal of this method is to produce a deeper understanding and a text-based predictive model to support a career
recommendation system. These steps are illustrated in the flowchart below.
Figure 2. Flowchart of the BERT model system
2.1.4 Application of Method
This study employs IndoBERT to classify Gen Z’s career preferences based on narrative texts from questionnaires.
IndoBERT was chosen because it is specifically designed for the Indonesian language and trained on a large Indonesian
language corpus, allowing it to understand the structure, context, and local vocabulary more accurately than the original
BERT model, which is based on English[17]. After going through preprocessing stages such as data cleaning,
normalization, stopword removal, and stemming, the data is trained and evaluated using metrics such as accuracy,
Copyright © 2025, Silvia Ramadhani. This is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. Page 100
The IJICS (International Journal of Informatics and Computer Science)
Vol 9 No 2, July 2025, Page 98-108
ISSN 2548-8384 (online), ISSN 2548-8449 (print)
Available Online at [Link]
DOI 10.30865/ijics.v9i2 .8917
precision, recall, and F1-score. This method demonstrates the effectiveness of natural language processing (NLP) in
deeply and contextually understanding and mapping Gen Z’s career preferences.
2.1.5 Testing Method
The method testing section aims to evaluate the performance of the IndoBERT model after being trained with
preprocessed questionnaire data. The dataset is divided into training and testing sets with a ratio of 80:20[18]. The model
is trained to recognize language patterns in each career category, then tested with new data to measure its ability to
generalize.
2.1.6 Results
The results section shows that the IndoBERT model successfully classified Gen Z's career preferences based on
questionnaire responses into four main categories: software engineer, content creator, digital marketing, and entrepreneur.
These results illustrate the distribution of career interests among respondents, such as the high level of interest in becoming
a content creator.
2.1.7 Evaluation
The evaluation stage aims to measure the performance of the IndoBERT model in classifying questionnaire text
responses based on the correct career labels. The evaluation uses metrics such as accuracy, precision, recall, and F1-score
to assess the model’s correctness, sensitivity, and balanced performance in recognizing patterns in the test data. This
evaluation ensures the model’s reliability in understanding Gen Z’s career preferences from narrative texts.
3. RESULT AND DISCUSSION
3.1 Data Collection
The data for this study was obtained through open-ended questionnaires distributed between May 27 and June 18,
2025, with a total of 602 respondents. The responses were converted into .csv format and processed using Python.
Preprocessing steps included normalization (removing punctuation, converting all text to lowercase, and removing
stopwords), followed by tokenization using the IndoBERT tokenizer. The labeled data was then split using the K-Fold
Cross Validation technique to ensure fair model evaluation. Sentiment classification was performed using the IndoBERT
model into two categories: positive and negative. Additionally, respondents’ career choices were analyzed to identify
career trends among Generation Z. The following is the sentiment label distribution of the entire dataset used.
Figure 3. Data distribution
3.2 Preprocessing Data
The data collected from respondents must go through preprocessing stages before it can be analyzed. The purpose of
preprocessing is to transform the initially unstructured and noisy dataset into clean data that is ready for processing. The
following is a flow that outlines the system stages during the data preprocessing process, which includes Data Cleaning,
Normalization, Stopword Removal, and Tokenization (IndoBERT).
a. Data Cleaning
Copyright © 2025, Silvia Ramadhani. This is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. Page 101
The IJICS (International Journal of Informatics and Computer Science)
Vol 9 No 2, July 2025, Page 98-108
ISSN 2548-8384 (online), ISSN 2548-8449 (print)
Available Online at [Link]
DOI 10.30865/ijics.v9i2 .8917
The cleaning stage aims to remove unnecessary elements from the data, such as numbers, symbols, emojis, links,
and punctuation marks. This process involves detecting these characters in each line of the data and removing them
to produce cleaner, more analyzable text. The table below shows a Sample of the Data Cleaning Process.
Table 1. Sample of Data Cleaning Process
Initial Sentiment Data Cleaning Text
sangat menarik, zaman sekarang semuanya sangat menarik, zaman sekarang semuanya
serba digital, jadi apa apa menggunakan AI, AI serba digital, jadi apa apa menggunakan ai, ai
termasuk digital yang bisa membantu kita termasuk digital yang bisa membantu kita
mencari informasi yang lebih dalam tentang mencari informasi yang lebih dalam tentang
pekerjaan mungkin mengikuti zaman karena pekerjaan mungkin mengikuti zaman karena
dengan menjadi content creator saya bisa dengan menjadi content creator saya bisa
menghasilkan uang dan menjadi seleb asek💃 menghasilkan uang dan menjadi seleb asek
banyaknya syarat" loker pada jaman sekarang banyaknya syarat" loker pada jaman sekarang
sehingga sulit untuk mendapatkan pekerjaan sehingga sulit untuk mendapatkan pekerjaan
media sosial kerjany fleksibel tidak di bawah media sosial kerjany fleksibel tidak di bawah
tekanan bos tekanan bos
Wah keren, sangat cerdas Karena fomo, gen z wah keren, sangat cerdas karena fomo, gen z ini
ini banyak kemauan nya jadi pasti mereka milih banyak kemauan nya jadi pasti mereka milih
karir yg kerja santai tapi penghasilan nya karir yg kerja santai tapi penghasilan nya
banyak Karena ini serba digital, sudah banyak karena ini serba digital, sudah dipastikan
dipastikan posisi ini cocok buat saya posisi ini cocok buat saya
perkembangan dunia kerja digital sekarang perkembangan dunia kerja digital sekarang
sangat pesat, lebih fleksibel, dan berbasis sangat pesat, lebih fleksibel, dan berbasis
teknologi gaji dan hari libur bisa kerja dr rumah teknologi gaji dan hari libur bisa kerja dr rumah
Sangat baik, mempermudah pencari kerja dan sangat baik, mempermudah pencari kerja dan
penyedia Sesuai minat dan hobi Pekerjaan yg penyedia sesuai minat dan hobi pekerjaan yg
saya sukai saya sukai
b. Normalization
After the data cleaning process, the next step is normalization, which involves removing symbols, punctuation
marks, and standardizing words. The table below shows a Sample of Normalization Process
Table 2. Sample of Normalization Process
Data Cleaning Text Normalized Text
sangat menarik, zaman sekarang semuanya serba sangat menarik zaman sekarang semuanya serba
digital, jadi apa apa menggunakan ai, ai digital jadi apa apa menggunakan ai ai termasuk
termasuk digital yang bisa membantu kita digital yang bisa membantu kita mencari
mencari informasi yang lebih dalam tentang informasi yang lebih dalam tentang pekerjaan
pekerjaan mungkin mengikuti zaman karena mungkin mengikuti zaman karena dengan
Copyright © 2025, Silvia Ramadhani. This is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. Page 102
The IJICS (International Journal of Informatics and Computer Science)
Vol 9 No 2, July 2025, Page 98-108
ISSN 2548-8384 (online), ISSN 2548-8449 (print)
Available Online at [Link]
DOI 10.30865/ijics.v9i2 .8917
dengan menjadi content creator saya bisa menjadi content creator saya bisa menghasilkan
menghasilkan uang dan menjadi seleb asek uang dan menjadi seleb asek
banyaknya syarat" loker pada jaman sekarang banyaknya syarat loker pada jaman sekarang
sehingga sulit untuk mendapatkan pekerjaan sehingga sulit untuk mendapatkan pekerjaan
media sosial kerjany fleksibel tidak di bawah media sosial kerjany fleksibel tidak di bawah
tekanan bos tekanan bos
wah keren, sangat cerdas karena fomo, gen z ini wah keren sangat cerdas karena fomo gen z ini
banyak kemauan nya jadi pasti mereka milih banyak kemauan nya jadi pasti mereka milih
karir yg kerja santai tapi penghasilan nya banyak karir yg kerja santai tapi penghasilan nya banyak
karena ini serba digital, sudah dipastikan posisi karena ini serba digital sudah dipastikan posisi
ini cocok buat saya ini cocok buat saya
perkembangan dunia kerja digital sekarang perkembangan dunia kerja digital sekarang
sangat pesat, lebih fleksibel, dan berbasis sangat pesat lebih fleksibel dan berbasis
teknologi gaji dan hari libur bisa kerja dr rumah teknologi gaji dan hari libur bisa kerja dr rumah
sangat baik, mempermudah pencari kerja dan sangat baik mempermudah pencari kerja dan
penyedia sesuai minat dan hobi pekerjaan yg penyedia sesuai minat dan hobi pekerjaan yg
saya sukai saya sukai
c. Stopword removal
Stopwords are common words such as “and,” “or,” “that,” and similar terms that frequently appear in text but
carry little meaning in sentiment analysis. Therefore, they are typically removed during data preprocessing to improve
efficiency and focus on more meaningful words. The table below shows Sample of Stopword Removal Process
Table 3. Sample of Stopword Removal Process
Normalized Text Stopword Removal Text
sangat menarik zaman sekarang semuanya serba sangat menarik zaman sekarang semuanya serba
digital jadi apa apa menggunakan ai ai termasuk digital jadi apa apa menggunakan ai ai termasuk
digital yang bisa membantu kita mencari digital bisa membantu mencari informasi lebih
informasi yang lebih dalam tentang pekerjaan dalam pekerjaan mungkin mengikuti zaman
mungkin mengikuti zaman karena dengan menjadi content creator bisa menghasilkan uang
menjadi content creator saya bisa menghasilkan menjadi seleb asek
uang dan menjadi seleb asek
banyaknya syarat loker pada jaman sekarang banyaknya syarat loker jaman sekarang sulit
sehingga sulit untuk mendapatkan pekerjaan mendapatkan pekerjaan media sosial kerjanya
media sosial kerjanya fleksibel tidak di bawah fleksibel tidak bawah tekanan bos
tekanan bos
wah keren sangat cerdas karena fomo gen z ini wah keren sangat cerdas fomo gen z banyak
banyak kemauan nya jadi pasti mereka milih kemauan jadi pasti milih karir yg kerja santai
karir yg kerja santai tapi penghasilan nya banyak tapi penghasilan banyak serba digital dipastikan
karena ini serba digital sudah dipastikan posisi posisi cocok buat
ini cocok buat saya
Copyright © 2025, Silvia Ramadhani. This is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. Page 103
The IJICS (International Journal of Informatics and Computer Science)
Vol 9 No 2, July 2025, Page 98-108
ISSN 2548-8384 (online), ISSN 2548-8449 (print)
Available Online at [Link]
DOI 10.30865/ijics.v9i2 .8917
perkembangan dunia kerja digital sekarang perkembangan dunia kerja digital sekarang
sangat pesat lebih fleksibel dan berbasis sangat pesat lebih fleksibel berbasis teknologi
teknologi gaji dan hari libur bisa kerja dr rumah gaji hari libur bisa kerja dr rumah
sangat baik mempermudah pencari kerja dan sangat baik mempermudah pencari kerja
penyedia sesuai minat dan hobi pekerjaan yg penyedia sesuai minat hobi pekerjaan yg sukai
saya sukai
d. Data Splitting
Data splitting in this study was carried out using the K-Fold Cross Validation technique. This method was chosen
to ensure that the model is fairly tested on unseen data and to provide more stable evaluation results that are not
dependent on a single data split. In this research, 5-Fold Cross Validation was used, where the entire dataset was
divided into five approximately equal parts.
Figure 4. Data Split Result
e. Tokenization (Indobert)
The model used is IndoBERT (indobenchmark/indobert-base-p1), a BERT-based model specifically trained on
Indonesian-language data. The IndoBERT tokenizer works by breaking down respondents' text into subword tokens
based on the vocabulary used by IndoBERT. These tokens are then converted into token IDs, which are integers
representing each token in the model's vocabulary. In the tokenization process, two datasets need to be tokenized:
the training data and the validation data. The table below shows Sample of the tokenization process (IndoBERT)
Table 4. Sample of the tokenization process (IndoBERT)
Teks tokenized_text input_ids attention_mask
ya sangat baik [CLS] ya sangat baik si mantap gen z 2 286 310 342 11111111
si mantap gen terkadang masi bnyk yg milih milih 356 7424 1449 11111111
z terkadang pekerjaan bidang sejak kuliah [SEP] [PAD] 653 3994 14025 11100000
masi bnyk yg [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] 21788 741 00000000
milih milih [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] 17693 17693 00000000
pekerjaan [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] 1367 1254 1098 00000000
bidang sejak [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] 3211 3 0 0 0 0 0 00000000
kuliah [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] 000000000 00000000
[PAD] [PAD] [PAD] [PAD] [PAD] [PAD] 000000000 00000000
[PAD] [PAD] [PAD] [PAD] [PAD] [PAD] 000000000 00000000
[PAD] [PAD] [PAD] [PAD] [PAD] [PAD] 000000000 00000000
[PAD] [PAD] [PAD] [PAD] [PAD] [PAD] 000000000 00000000
[PAD] [PAD] [PAD] [PAD] [PAD] [PAD] 000000000 00000000
[PAD] [PAD] [PAD] [PAD] [PAD] [PAD] 000000000 00000000
[PAD] [PAD] [PAD] [PAD] [PAD] [PAD] 000000000 00000000
[PAD] [PAD] [PAD] [PAD] [PAD] [PAD] 000000000 00000000
[PAD] [PAD] [PAD] [PAD] [PAD] [PAD] 000000000
[PAD] [PAD] [PAD] [PAD] [PAD] [PAD]
Copyright © 2025, Silvia Ramadhani. This is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. Page 104
The IJICS (International Journal of Informatics and Computer Science)
Vol 9 No 2, July 2025, Page 98-108
ISSN 2548-8384 (online), ISSN 2548-8449 (print)
Available Online at [Link]
DOI 10.30865/ijics.v9i2 .8917
[PAD] [PAD] [PAD] [PAD] [PAD] [PAD] 000000000
[PAD] [PAD] [PAD] [PAD] [PAD] [PAD] 00000
[PAD] [PAD] [PAD] [PAD] [PAD] [PAD]
persaingan [CLS] persaingan kerja makin ketat tapi 2 6283 494 2855 11111111
kerja makin makin banyak peluang buat punya skill 5983 469 2855 11111111
ketat tapi terutama skill bidang it kerja bisa 271 2882 968 11111111
makin banyak dikerjakan mana aja bidang teknologi 1121 7963 1580 11111111
peluang buat punya banyak peluang pilih software 7963 1254 1701 11111111
punya skill engineer suka logika cod ##ing rasanya 494 166 6565 11100000
terutama skill senang kalau bisa buat program bisa bantu 1152 1465 1254 00000000
bidang it kerja orang [SEP] [PAD] [PAD] [PAD] [PAD] 1429 1121 271 00000000
bisa [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] 2882 959 3389 00000000
dikerjakan [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] 21058 1506 7858 00000000
mana aja [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] 11114 55 3514 00000000
bidang [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] 3000 599 166 968 00000000
teknologi [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] 986 166 1621 232 00000000
punya banyak [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] 3000000000 00000000
peluang pilih [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] 0000000000 00000000
software [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] 0000000000 00000000
engineer suka [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] 0000000000
logika coding [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] 0000000000
rasanya [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] 0000000000
senang kalau [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] 0000000000
bisa buat [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] 0000000000
program bisa [PAD] [PAD] [PAD] 000000
bantu orang
3.3 Implementation of the BERT Model
In this study, the IndoBERT model (indobenchmark/indobert-base-p1) was used for sentiment classification of
Indonesian-language text. After preprocessing steps such as normalization, stopword removal, and tokenization, the data
were converted into token IDs and attention masks to be fed into the BERT for Sequence Classification model. The model
was trained for three epochs using the AdamW optimizer with a learning rate of 2e-5. Evaluation was performed using
K-Fold Cross Validation with metrics including accuracy, precision, recall, and F1-score.
3.4 Model Evaluation
The horizontal bar chart visualizes the career preferences of Gen Z respondents. The graph displays the number of
respondents on the horizontal axis and the types of careers on the vertical axis. The results show that Content Creator is
the most preferred career choice (183 respondents), followed by Entrepreneur (157), Digital Marketing (146), and
Software Engineer as the least chosen (115). These findings indicate that Gen Z tends to be more interested in careers that
support creativity and self-expression rather than technical professions.
Copyright © 2025, Silvia Ramadhani. This is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. Page 105
The IJICS (International Journal of Informatics and Computer Science)
Vol 9 No 2, July 2025, Page 98-108
ISSN 2548-8384 (online), ISSN 2548-8449 (print)
Available Online at [Link]
DOI 10.30865/ijics.v9i2 .8917
Figure 5. Career Choices Preferred by Gen Z
a. Confusion matrix
The confusion matrix shows the number of correct and incorrect predictions for both positive and negative
classes, allowing us to understand how well the model can distinguish between the two. This evaluation result
provides a comprehensive overview of the model’s ability to generalize to new data.
The following figure presents the results of the confusion matrix, which illustrates the number of correct and
incorrect predictions made by the model for each class, both positive and negative. This visualization helps evaluate
how accurately the model classifies the data.
Figure 6. Confusion Matrix
b. Accuracy
The accuracy score measures the proportion of correct predictions overall. It is calculated by dividing the number
of correct predictions by the total number of predictions. A high accuracy score indicates that the model performs
well overall in correctly classifying sentiments.
The image below presents the accuracy results achieved by the IndoBERT model in classifying the data. This
visualization provides an overview of how effectively the model can recognize patterns and make accurate
predictions based on the data that has undergone the training process.
Figure 7. Accuracy
58 + 4 62
𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = = × 100% = 51,24%
58 + 4 + 53 + 6 121
c. Precision
1. For the negative class
Copyright © 2025, Silvia Ramadhani. This is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. Page 106
The IJICS (International Journal of Informatics and Computer Science)
Vol 9 No 2, July 2025, Page 98-108
ISSN 2548-8384 (online), ISSN 2548-8449 (print)
Available Online at [Link]
DOI 10.30865/ijics.v9i2 .8917
4 4
𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 = = × 100% = 40%
4 + 6 10
2. For the Positive class
58 58
𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 = = × 100% = 52,25%
58 + 53 111
d. Recall
1. For the negative class
4 4
𝑅𝑒𝑐𝑎𝑙𝑙 = = × 100% = 7,02%
4 + 53 57
2. For the Positive class
58 58
𝑅𝑒𝑐𝑎𝑙𝑙 = = × 100% = 90,62%
58 + 6 64
e. F1 - Score
1. For the negative class
0,4 × 0,0702 0,0281
𝐹1 − 𝑆𝑐𝑜𝑟𝑒 = 2 × =2× × 100% = 11,97%
0,4 + 0,0702 0,4702
2. For the Positive class
0,5225 × 0,9062 0,4739
𝐹1 − 𝑆𝑐𝑜𝑟𝑒 = 2 × =2× × 100% = 66,34%
0,5225 + 0,9062 1,4287
3.5 Implementation
This study utilizes the BERT (Bidirectional Encoder Representations from Transformers) model to analyze textual
data from questionnaires regarding Generation Z's career preferences in the digital era. After undergoing normalization,
text cleaning, and stopword removal, the data is tokenized using the IndoBERT tokenizer and then split into training and
validation sets. The model is trained over several epochs while monitoring training loss and validation accuracy.
Evaluation is conducted using accuracy metrics, a confusion matrix, and performance visualizations. The results are used
to identify sentiment patterns and career trends among Gen Z, providing deeper insights into the factors influencing their
career decisions in the digital age.
4. CONCLUSION
Based on the final results of the problem-solving process in this study, several conclusions can be drawn as follows.
First, from a total of 601 questionnaire responses regarding Gen Z’s career preferences, it was found that 57.85%
expressed positive sentiment, while 42.15% reflected negative sentiment. Second, the implementation of the BERT model
to classify Gen Z’s career choices successfully mapped respondents into four predefined categories: content creator,
entrepreneur, digital marketing, and software engineering. The prediction results showed that the majority of respondents
preferred careers as content creators, followed by entrepreneurs, then digital marketers, with software engineering being
the least chosen. These findings indicate a strong interest among Generation Z in creative, independent, and digital fields.
Third, the evaluation of the IndoBERT model during training, visualized through graphs and epoch-based metrics,
showed a decreasing trend in training loss from epoch 1 to 3 indicating improved learning on the training data. However,
validation accuracy remained stagnant between 47% and 49%, suggesting that the model struggled to generalize to the
validation set despite a sharp drop in training loss from 0.32 to 0.11. Lastly, the BERT model used in this study achieved
a fairly good accuracy rate of 51.24% in classifying career choices based on opinion texts. The evaluation results showed
varied precision, recall, and F1-scores across categories, with generally satisfactory performance. The model performed
well on categories with larger datasets, while its performance was relatively lower for categories with fewer data samples.
REFERENCES
[1] S. D. Saraswati, N. Alifia, and G. B. Setyawan, “YUME : Journal of Management Pengaruh Persepsi Generasi Z Terhadap
Pilihan Karir Pada Mahasiswa Universitas Karya Husada,” vol. 8, no. 2, pp. 1691–1699, 2025.
[2] M. Alimuddin and P. Poddala, “Prospek Digital Marketing Untuk Generasi Muda Dalam Berwirausaha,” J. Career Dev., vol.
1, no. 1, pp. 54–70, 2023.
[3] Eka Nurkhasannah and Shafa Anggita Putri, “Pengaruh Persepsi Generasi Z Terhadap Pilihan Karir pada Mahasiswa Program
Studi Manajemen Universitas Pamulang,” J. Manag. Creat. Bus., vol. 3, no. 1, pp. 251–260, 2025, doi:
10.30640/jmcbus.v3i1.3703.
Copyright © 2025, Silvia Ramadhani. This is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. Page 107
The IJICS (International Journal of Informatics and Computer Science)
Vol 9 No 2, July 2025, Page 98-108
ISSN 2548-8384 (online), ISSN 2548-8449 (print)
Available Online at [Link]
DOI 10.30865/ijics.v9i2 .8917
[4] D. Mawabagja, A. Kurniawan, R. Claudia, S. Islami, and R. Fadjri, “Strategi Adaptasi Manajemen Puncak dalam Menghadapi
Transformasi Tenaga Kerja Generasi Z Pendahuluan Transformasi tenaga kerja yang terjadi saat ini dipengaruhi oleh
masuknya generasi Z,” vol. 03, no. 03, pp. 132–142, 2025.
[5] Z. W. Farida and N. Rochmawati, “Analisis Sentimen Masyarakat terhadap Fenomena Childfree Menggunakan Metode Long
Short Term Memory dan Bidirectional Encoder Representations from Transformers di Twitter,” J. Informatics Comput. Sci.,
vol. 5, no. 03, pp. 369–376, 2024, doi: 10.26740/jinacs.v5n03.p369-376.
[6] R. M. R. W. P. K. Atmaja and W. Yustanti, “Analisis Sentimen Customer Review Aplikasi Ruang Guru dengan Metode BERT
(Bidirectional Encoder Representations from Transformers),” Jeisbi, vol. 02, no. 03, p. 2021, 2021.
[7] S. Prasetyo and T. Dewayanto, “Penerapan Machine Learning, Deep Learning, Dan Data Mining Dalam Deteksi Kecurangan
Laporan Keuangan-a Systematic Literature Review,” Diponegoro J. Account., vol. 13, no. 3, pp. 1–12, 2024, [Online].
Available: [Link]
[8] R. Kurniawan, P. B. Wintoro, Y. Mulyani, and M. Komarudin, “Implementasi Arsitektur Xception Pada Model Machine
Learning Klasifikasi Sampah Anorganik,” J. Inform. dan Tek. Elektro Terap., vol. 11, no. 2, pp. 233–236, 2023, doi:
10.23960/jitet.v11i2.3034.
[9] I. C. Azhari and T. Haryanto, “Modeling Of Hyperparameter Tuned RNN-LSTM and Deep Learning For Garlic Price
Forecasting In Indonesia,” J. Informatics Telecommun. Eng., vol. 7, no. 2, pp. 502–513, 2024, doi: 10.31289/jite.v7i2.10714.
[10] P. Hafidzah, S. Maryani, B. Y. Ihsani, E. Erwin, and A. K. Niswariyana, “Penerapan Deep Learning dalam Menganalisis
Sentimen di Media Sosial,” Semin. Nas. Paedagoria, vol. 4, pp. 328–339, 2024.
[11] Vidya Chandradev, I Made Agus Dwi Suarjaya, and I Putu Agung Bayupati, “Analisis Sentimen Review Hotel Menggunakan
Metode Deep Learning BERT,” J. Buana Inform., vol. 14, no. 02, pp. 107–116, 2023, doi: 10.24002/jbi.v14i02.7244.
[12] M. Furqan, S. Sriani, and M. N. Shidqi, “Chatbot Telegram Menggunakan Natural Language Processing,” Walisongo J. Inf.
Technol., vol. 5, no. 1, pp. 15–26, 2023, doi: 10.21580/wjit.2023.5.1.14793.
[13] W. Alan, “Klasifikasi Emosi pada Teks Menggunakan Deep Learning,” Tugas Akhir di Univ. Islam Indones., vol. 6, no. 1,
2021.
[14] A. Ripa’i, F. Santoso, and F. Lazim, “Deteksi Berita Hoax dengan Perbandingan Website Menggunakan Pendekatan Deep
Learning Algoritma BERT,” G-Tech J. Teknol. Terap., vol. 8, no. 3, pp. 1749–1758, 2024, doi: 10.33379/gtech.v8i3.4541.
[15] Y. A. Singgalen, “IndoBERT-Based Sentiment Analysis for Understanding Hotel Guests ’ IndoBERT-Based Sentiment
Analysis for Understanding Hotel Guests ’ Preferences,” no. March, 2025, doi: 10.47065/josyc.v6i2.6864.
[16] E. Mundika, G. H. Martono, and R. Rismayati, “Sentiment Analysis of Service and Facility Satisfaction at Computer Lab of
Universitas Bumigora Using Indobert,” vol. 5, no. 2, pp. 475–484, 2025, doi: 10.30811/jaise.v5i2.6798.
[17] G. T. Fadilah, L. Muflikhah, and R. S. Perdana, “Analisis Sentimen Produk Hijab Pada E-Commerce Tokopedia Menggunakan
Algoritma Support Vector Machine Dan Indobert Embedding,” vol. 9, no. 2, pp. 2548–964, 2025, [Online]. Available: [Link]
[Link]
[18] A. S. Rizky and E. Y. Hidayat, “Emotion Classification in Indonesian Text Using IndoBERT,” Comput. Eng. Appl., vol. 14,
no. 1, pp. 2252–4274, 2025.
.
Copyright © 2025, Silvia Ramadhani. This is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. Page 108