Comparative Analysis of IndoBERT and LSTM for MultiLabel Text Classification of Indonesian Motivation Letter - Deep Learning Benchmark

Sep 11, 2026

Evaluation of motivation letters is a pivotal step in vocational student admissions in Indonesia, yet manual scoring often introduces subjectivity and inconsistency that undermines fairness. This study compares two deep learning models, IndoBERT and Long Short-Term Memory (LSTM), for multi-label text classification of Indonesian motivation letters. Using 676 letters across nine predefined categories, the research assesses classification performance and reports stronger results for IndoBERT (F1-score 81%) versus LSTM (76%). The findings support automation and establish a benchmark for future work.

LanguageEnglish

CategoryPresentations

FormatPDF

Pages10

Reading Time4 min

AccessFree to download

Abstract

Introduction

Method

Flow Chart

Data Preparation

Comparative Analysis of IndoBERT and LSTM for MultiLabel Text Classification of Indonesian Motivation Letter Yosep Setiawan1, Lili Ayu Wulandhari2 1,2Department of Computer Science, Binus University Jakarta, Indonesia Article Info ABSTRACT | Article history:
Received November 18, 2024 Revised February 06, 2025 Accepted February 06, 2025 Published August 17, 2025 | The evaluation of motivation letters is a crucial step in the student admission process for one of vocational institutions in Indonesia. However, the current manual assessment method is prone to subjectivity and inconsistency, making it less reliable for fair studentselection. This research presents a comparative analysis of two deep learning models, IndoBERT and Long Short-Term Memory (LSTM), for multi-label text classification of motivation letters written in Indonesian. Using a dataset of 676 motivation letters labeled with nine predefined categories, we evaluate the models based on their classification performance. The results indicate that IndoBERT outperforms LSTM, achieving an F1-score of 81%, compared to 76% for LSTM. This research provides insights into the effectiveness of IndoBERT for multi-label classification tasks in the Indonesian language and serves as a benchmark for future research in automating motivation letter evaluations. | | --- | --- | | Keywords:
Comparative analysis IndoBERT
LSTM
Motivation letter
Multi-label text classification | | Corresponding Author: Yosep Setiawan, Department of Computer Science, Binus University Jakarta, Indonesia Jl. Kebon Jeruk Raya No. 27, Kebon Jeruk Jakarta Barat 11530, Indonesia [Email: yosep.setiawan001@binus.ac.id](Email: yosep.setiawan001@binus.ac.id) 1. INTRODUCTION The evaluation of motivation letters is a crucial step in the student selection process at vocational institutions in Indonesia. These letters reflect applicants’ aspirations, motivations, and suitability for their chosen research programs, but manual evaluation is prone to subjectivity and inconsistency, raising concerns about fairness and reliability. To improve objectivity and efficiency, this research proposes a multi-label text classification approach capable of assessing multiple aspects of motivation letters simultaneously, such as motivation, achievements, and instructional suitability. Instead of assigning only one label per document, multiple labels capture the different evaluation dimensions present in each letter. Advances in deep learning have introduced models such as BERT and LSTM, which excel in handling complex text classification tasks. Prior studies have shown BERT’s superior performance in document classification compared to CNN, LSTM, and SVM models [1]–[4] . In the Indonesian NLP context, IndoBERT, which is BERT pretrained on large-scale Indonesian text has demonstrated strong results in various multi-label classification tasks, including customer review analysis, toxic comment detection, and sentiment classification [5]–[8]. These works indicate IndoBERT’s capacity to capture rich contextual information across domains. LSTM-based architectures, on the other hand, are valued for their ability to capture sequential dependencies in text. They have been successfully applied in multi-label classification across languages and domains [9]–[11], and continue to be competitive, especially when enhanced with mechanisms such as attention layers. However, no prior research has directly compared IndoBERT and LSTM for multi-label text classification of Indonesian motivation letters. Addressing this gap, our research provides a direct performance comparison and establishes a benchmark for future work in automated motivation letter evaluation, contributing to both educational technology and the broader field of Indonesian NLP. 2. METHOD In this section, we outline the research methodology of the proposed approach, offering a detailed discussion of each step. We introduce a deep learning mode

FAQs

Why is multi-label classification used for motivation letters instead of single-label scoring?

Motivation letters contain multiple evaluation dimensions, such as motivation, achievements, and suitability. Multi-label classification assigns several relevant labels to capture these aspects simultaneously.

What dataset and labeling scheme are used in the study?

The study uses 676 motivation letter files submitted in 2022 and 2023, labeled into nine predefined categories for multi-label classification.

How does IndoBERT performance compare with LSTM?

IndoBERT outperforms LSTM, reaching an F1-score of 81% compared to 76% for LSTM on the multi-label task.

Sep 11, 2026

Upload by Stanford