Jul 2026· JURNAL ILMIAH SAINS TEKNOLOGI DAN INFORMASI· Vol 4, pp. 409-422· 0 citations· 19 references
TL;DR
Overall, the study confirms that Bi-LSTM is a suitable deep learning approach for sentiment classification of application reviews and offers meaningful insights that can support Duolingo developers in evaluating user opinions and enhancing application quality.
Abstract
Reviews posted by users on the Google Play Store provide valuable feedback that can be utilized to measure user satisfaction and assess the quality of mobile applications. However, the growing volume of reviews makes manual evaluation increasingly impractical, highlighting the need for automated sentiment analysis techniques. This research proposes the use of the Bidirectional Long Short-Term Memory (Bi-LSTM) algorithm to classify the sentiment of Indonesian-language reviews for the Duolingo application. The review dataset was obtained through web scraping from the Google Play Store and underwent several preprocessing steps, including case folding, text cleaning, word normalization, tokenization, stopword removal, and stemming. After preprocessing, the data were divided into 80% training data and 20% testing data for model development and performance evaluation. The effectiveness of the model was measured using accuracy, precision, recall, and F1-score. The experimental results yielded an accuracy of 94.05%, precision of 88.46%, recall of 92.15%, and an F1-score of 91.17%. These findings indicate that the Bi-LSTM model is capable of capturing sentiment patterns with a high level of reliability, although its ability to classify negative reviews is still influenced by the imbalance between sentiment classes. Overall, the study confirms that Bi-LSTM is a suitable deep learning approach for sentiment classification of application reviews and offers meaningful insights that can support Duolingo developers in evaluating user opinions and enhancing application quality.
Shopee is one of the most widely used e-commerce applications in Indonesia, and the reviews written by its users on the Google Play Store contain valuable information about service quality, application performance, and customer satisfaction. This study aims to classify the sentiment of Indonesian-language reviews of the Shopee application using the Random Forest algorithm. A total of 5,000 reviews were collected through web scraping, labeled based on user ratings, and processed through cleaning, case folding, slang-word normalization, tokenization, stopword removal, and stemming. Feature extraction was performed using Term Frequency-Inverse Document Frequency (TF-IDF), and the Synthetic Minority Over-sampling Technique (SMOTE) was applied to handle class imbalance in the training data. The experimental results show that the best Random Forest model, with 200 trees, achieves an accuracy of 89.34%, a precision of 89.40%, a recall of 88.30%, and an F1-score of 88.75%, outperforming Naive Bayes, Support Vector Machine, and K-Nearest Neighbor as comparison models. The analysis of feature importance shows that positive sentiment is dominated by words related to delivery speed and price, while negative sentiment is dominated by complaints about system errors, sellers, and refund processes. These findings can be used by application managers to prioritize service improvements.
Sopi Sapriadi, Rahmatia Wulan Dari, Y. Eirlangga· Jurnal Teknologi Dan Sistem...· 0 citations
On the Google Play Store platform, there is a review section for each app containing users' opinions and experiences when using an app. These user reviews can be used as a basis for evaluating the quality of an app’s service; however, the large number of reviews makes it difficult for developers to analyze them manually. Threads, a social media app that provides a means of communication and entertainment created by Meta, has received many reviews from its users. This study analyzes user sentiment toward the Threads app through review classification to identify positive and negative opinions. A total of 1,808 user review data for the Threads app was collected as research data, consisting of 961 positive data and 847 negative data. Sentiment labeling will be done using the AI Copilot tool. The data will go through several stages, including data selection, data cleaning, data normalization, and word weighting using the TF-IDF method before performing data mining using the Support Vector Machine (SVM) algorithm. The test results showed that the model built was able to classify sentiment with an accuracy rate of 93.03% on an 80:20 train-test data split using the rbf kernel. In the Word Cloud, positive sentiments were dominated by words related to users’ appreciation for the Threads app services, while negative sentiments were dominated by users’ complaints about the suspension system and features in the Threads app. These research results indicate that the method used is capable of identifying user opinions and can be used as a basis for evaluating improvements in app service quality.
The rapid growth of e-commerce applications has increased the number of user reviews that reflect public opinion on service quality and user experience. However, many previous studies rely only on rating-based sentiment analysis and do not utilize positive and negative sentiment lexicons, resulting in limited insight into public opinion. This study aims to analyze public opinion toward the Lazada application using a data mining–based sentiment analysis approach by combining user ratings and sentiment lexicons in the labeling process. The data set consists of Lazada user reviews collected from digital platforms. The research process includes data collection, data cleaning, and text preprocessing, such as text normalization, removal of emojis and symbols, and stopword elimination. Feature extraction is performed using the Term Frequency–Inverse Document Frequency (TF-IDF) method. Sentiment labeling is conducted by integrating rating scores and lexicon analysis, where reviews with ratings of 4–5 or dominant positive words are classified as positive, while reviews with ratings of 1–2 or dominant negative words are classified as negative; neutral reviews are excluded. Sentiment classification is carried out using Naive Bayes and Random Forest algorithms. Model performance is evaluated using accuracy, precision, recall, and F1-score. The results show that both models perform well, with Random Forest achieving better performance than Naive Bayes. This study provides useful insights for improving e-commerce service quality based on user feedback.
Mohammad Rizal, Nur Azizah, Firman Jaya· Jurnal Informatika Dan Tekon...· 0 citations
The rapid growth of PLN Mobile users has generated massive volumes of review data that are impractical to analyze manually, creating an urgent need for automated customer satisfaction monitoring. This study aims to develop an accurate sentiment classification model for PLN Mobile user reviews from the Google Play Store using the Naive Bayes algorithm. A dataset of 171,000 reviews was collected via web scraping, labeled into positive (rating 5) and negative (rating 1–2) classes, and balanced using an undersampling technique. Indonesian text preprocessing included text cleaning, repeated character removal, slang word normalization, negation handling, intensifier handling, stopword removal, and stemming using the Sastrawi library. Feature representation employed the Bag of Words method with CountVectorizer considering unigrams and bigrams. The Multinomial Naive Bayes model achieved an accuracy of 95.31% on the test set, with precision of 93% for the negative class and 97% for the positive class, recall of 98% for negative and 93% for positive, and an F1-score of 95% for both classes. These findings confirm that Naive Bayes with comprehensive Indonesian text preprocessing is effective for automated customer satisfaction monitoring in digital electricity services.
Ni Kadek, Anggita Pradnya Dewi, N. Nyoman et al.· Jurnal Teknologi Dan Sistem...· 0 citations
