Topic Modeling and Sentiment Analysis on News Headlines Using BERTopic and IndoBERT Models

YouTube Sentiment Analysis on Felt Earthquake News Using LSTM and IndoBERT

This study analyzes public sentiment in YouTube comments on felt-earthquake news in Indonesia and compares a bidirectional Long Short-Term Memory implementation (LSTM) with IndoBERT. An experimental quantitative design was used. Comments were collected through the YouTube Data API v3 using official earthquake-event references from the Indonesian Agency for Meteorology, Climatology, and Geophysics (BMKG) and keyword-based video searches covering January 2021 to August 2025. The acquisition stage produced 51,870 comments from 1,626 unique videos. Data were processed through duplicate removal, text cleaning, case folding, slang normalization, tokenization, and stopword removal, resulting in 49,041 clean comments. Positive, negative, and neutral labels were assigned by aggregating word-polarity scores from the Indonesian Sentiment Lexicon (InSet), which served as weak supervision. Stratified sampling divided the dataset into 80% training data and 20% testing data. The LSTM model used a 100-dimensional embedding, a 64-unit bidirectional LSTM layer, global max pooling, and early stopping; IndoBERT was fine-tuned from indobenchmark/indobert-base-p2 for four epochs. Performance was assessed with accuracy, macro precision, macro recall, macro F1-score, and confusion matrices. Positive sentiment accounted for 41.2% of the corpus, negative sentiment for 35.5%, and neutral sentiment for 23.2%. IndoBERT achieved 91.50% accuracy and a 91.03% macro F1-score, outperforming LSTM at 90.91% accuracy and a 90.34% macro F1-score. IndoBERT provided the strongest contextual classification, while LSTM remained a competitive and substantially lighter option for resource-constrained monitoring

Sentiment Analysis of User Reviews of Purwokerto Station Using TF-IDF and the Random Forest Method

User reviews on Google Maps serve as a valuable source of information for evaluating the service quality of public facilities, including train stations. This study aims to analyze the sentiment of user reviews for Purwokerto Station using the Random Forest algorithm and to identify dominant Bigrams within each sentiment class to provide insights for station management. The dataset was obtained by scraping Google Maps, yielding 2,232 reviews after data cleaning. The research process involved sentiment labeling based on ratings; text preprocessing (cleaning, case folding, tokenization, non-standard word normalization, stopword removal, and single-token deduplication); and feature extraction using Bigram-based Term Frequency–Inverse Document Frequency (TF–IDF). The dataset was split into 80% training data and 20% testing data, then classified using the Random Forest algorithm. Results indicate that the model achieved 90.16% accuracy on the test data. However, evaluation revealed that while the model excelled at recognizing the positive class, it struggled to identify negative and neutral classes; this was evidenced by a Macro F1-score of 0.32 and precision, recall, and F1-score values ​​of 0.00 for both the negative and neutral classes. This outcome was driven by an imbalanced data distribution, causing the model to lean toward predicting reviews as positive. Bigram analysis showed that the phrase "clean and comfortable"  dominated positive sentiment, whereas phrases such as "waiting room" and "ticket counter staff" more frequently in negative sentiment. ABSTRAK (Bahasa Indonesia) Ulasan pengguna pada platform Google Maps menjadi salah satu sumber informasi yang dapat dimanfaatkan untuk mengevaluasi kualitas pelayanan suatu fasilitas publik, termasuk stasiun kereta api. Penelitian ini bertujuan menganalisis sentimen ulasan pengguna Stasiun Purwokerto menggunakan algoritma Random Forest serta mengidentifikasi Bigram yang dominan pada setiap kelas sentimen sebagai bahan evaluasi bagi pengelola stasiun. Dataset diperoleh melalui scraping Google Maps dan menghasilkan 2.232 ulasan setelah melalui proses pembersihan data. Tahapan penelitian meliputi pelabelan sentimen berdasarkan rating, preprocessing text yang terdiri atas cleaning, case folding, tokenizing, normalisasi kata tidak baku, stopword removal, dan duplikasi token tunggal, kemudian dilanjutkan dengan ekstraksi fitur menggunakan Term Frequency–Inverse Document Frequency (TF–IDF) berbasis Bigram. Dataset dibagi menjadi data pelatihan sebesar 80% dan data pengujian sebesar 20%, kemudian diklasifikasikan menggunakan algoritma Random Forest. Hasil penelitian menunjukkan bahwa model memperoleh akurasi sebesar 90,16% pada data pengujian. Namun, hasil evaluasi juga menunjukkan bahwa model memiliki kemampuan yang sangat baik dalam mengenali kelas positif, tetapi belum mampu mengidentifikasi kelas negatif dan netral, yang ditunjukkan oleh nilai Macro F1-score sebesar 0,32 serta nilai precision, recall, dan F1-score sebesar 0,00 pada kedua kelas tersebut. Kondisi tersebut dipengaruhi oleh distribusi data yang tidak seimbang sehingga model cenderung memprediksi ulasan ke kelas positif. Analisis Bigram menunjukkan bahwa frasa bersih nyaman mendominasi sentimen positif, sedangkan frasa seperti ruang tunggu dan petugas loket lebih banyak muncul pada sentimen negatif. Penelitian ini menunjukkan bahwa representasi TF–IDF Bigram mampu menangkap konteks frasa yang membedakan setiap kategori sentimen serta memberikan informasi mengenai aspek pelayanan yang paling sering diapresiasi maupun dikeluhkan oleh pengguna.  

A short text sentiment analysis method integrating BiGRU and additive attention

In the context of the digital economy, e-commerce and social media have generated massive amounts of short Chinese web texts, making the accurate extraction of sentiment information a critical requirement for market analysis and public opinion monitoring. Short texts are characterized by fragmented information expression, strong context dependency, and sparse features, posing unique challenges for sentiment analysis. To address the shortcomings of traditional methods in terms of semantic understanding depth and scenario adaptability, this study proposes a short-text sentiment analysis method (AddAttn-BiGRU) that integrates BiGRU with Additive Attention. This method employs the BiGRU model to extract semantic information from text context and incorporates a custom Additive Attention mechanism, enabling the model to autonomously learn the emotional importance of different words in short texts, assign differentiated weights, and focus on core emotional words. Validation experiments were conducted on the publicly available Weibo Sentiment 100k dataset. The results showed an accuracy of 97.21% and an F1 score of 97.66%, validating the effectiveness of the model.

Constructing high-frequency and thematic economic sentiment indicators from online news articles: applications in the Philippine context 1

Comparison of Machine Learning Algorithms for Sentiment Analysis of Trans Jogja on Social Media

The social media platform X has become an important channel for the public to express opinions and share experiences regarding public services, including Trans Jogja. User-generated content from this platform provides valuable insights into public perceptions of service quality. However, because these data consist of unstructured text, sentiment classification techniques based on machine learning are required to analyze them effectively. This study aims to compare the performance of several machine learning algorithms for sentiment classification, including Naïve Bayes, Support Vector Machine (SVM), Random Forest, Neural Network, Logistic Regression, and Decision Tree, in classifying user sentiment toward Trans Jogja on the X platform. Data were collected through a web crawling process using Tweet Harvest with keywords related to Trans Jogja, covering the period from January 1, 2025, to June 10, 2026, resulting in a dataset of 3,035 tweets. The preprocessing stage included data cleaning, case folding, tokenization, normalization, stopword removal, and stemming. Text representation was performed using the Term Frequency–Inverse Document Frequency (TF-IDF) method. The dataset was then divided into training and testing sets using five train–test split ratios: 90:10, 85:15, 80:20, 75:25, and 70:30. Model performance was evaluated using a confusion matrix and the corresponding accuracy, precision, recall, and F1-score metrics. The experimental results demonstrate that the Support Vector Machine (SVM) consistently outperformed the other algorithms across different data split ratios. At the 85:15 train–test split, the SVM achieved an accuracy of 91%, precision of 91%, recall of 91%, and an F1-score of 91%, indicating that it is the most effective algorithm for sentiment classification of Trans Jogja users on the X platform.

A Multi-Algorithm Model Selector for Sentiment Classification in Public Evaluations of Government Performance Using GovBERTic

GovBERTic (Government BERTopic), a reproducible computational pipeline that operationalizes public opinion analysis by linking semantic topic discovery, multi-algorithm sentiment model selection, and topic-level sentiment aggregation within a single workflow, is developed.