Aspect-Based Sentiment Analysis with LDA and IndoBERT Algorithm on Mental Health App: Riliv
Indonesia's mental health crisis in 2024 is increasing along with the high growth of internet users. Thus, high internet usage provides an opportunity for mobile applications including Riliv, a popular mental health application in Indonesia to become the most complete solution for overthinking,...
Saved in:
| Main Authors: | , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Politeknik Negeri Batam
2025-03-01
|
| Series: | Journal of Applied Informatics and Computing |
| Subjects: | |
| Online Access: | https://jurnal.polibatam.ac.id/index.php/JAIC/article/view/8958 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Indonesia's mental health crisis in 2024 is increasing along with the high growth of internet users. Thus, high internet usage provides an opportunity for mobile applications including Riliv, a popular mental health application in Indonesia to become the most complete solution for overthinking, anxiety, and depression. This research aims to analyze the sentiment correlation of aspects based on App Store and Play Store reviews through scraping to effectively expose Riliv’s user satisfaction knowledge to developers using topic labeling with Latent Dirichlet Allocation (LDA) and sentiment labeling using Bidirectional Encoder Representations from Transformers (BERT) indobenchmark/indobert-base-p1 model on Aspect-Based Sentiment Analysis (ABSA). This study used 3068 reviews from September 2015 to December 2024. The main results obtained were 1) Identified the sentiment that positive is highest in 2020, neutral is highest in 2020, and negative is highest in 2018. 2) Identified 4 main aspects of the Riliv application: Access Support, Counseling Services, Meditation Features, and User Interface with LDA. 3) The majority distribution was negative on User Interface, neutral on Counseling Services, and positive on Meditation Features. 4) The effectiveness of IndoBERT compared to the non-transformer baseline algorithm. 5) The most optimal results were obtained with 70% training, 10% validation, and 20% testing, resulting in 95% accuracy. |
|---|---|
| ISSN: | 2548-6861 |