Journal of Digital Society

Emotion Detection in Railway Complaints Using Deep Learning and Transformer Models: A Data Mining Approach to Analyzing Public Sentiment on Twitter

Jeffri Prayitno Bangkit Saputra, Aayush Kumar — Mon, 02 Jun 2025 00:00:00 +0700

In the era of digital communication, social media platforms like Twitter have become pivotal channels for passengers to express their experiences and grievances regarding public transportation services. Traditional sentiment analysis methods, which broadly classify text into positive, negative, or neutral categories, often fail to capture the complex emotional nuances embedded in such complaints. This study aims to bridge this gap by leveraging the Bidirectional Encoder Representations from Transformers (BERT) model to perform fine-grained emotion detection on railway complaint tweets. Using a dataset of 1,366 tweets labeled with multiple sentiment categories, we preprocess the data through comprehensive cleaning techniques and extract textual features using Term Frequency-Inverse Document Frequency (TF-IDF) vectorization. A logistic regression classifier was trained on these features as a baseline, achieving an overall accuracy of 74.45%, demonstrating the viability of text-based emotion classification in this domain. The analysis further identified key linguistic features associated with different emotional categories, such as frustration linked to hygiene and delay complaints, and satisfaction reflected in polite expressions. By correlating detected emotional intensity with complaint severity, the study revealed that heightened emotions, especially anger and urgency, often signal more critical service failures requiring prompt attention. These insights suggest that incorporating emotion detection into complaint management can significantly enhance railway service responsiveness and customer satisfaction. This research contributes to the growing field of emotion mining in digital society studies by applying advanced natural language processing techniques to a specific and socially impactful domain—railway services. The findings advocate for a more empathetic approach to handling customer feedback, moving beyond surface-level sentiment analysis to understand passengers’ emotional experiences deeply. Future work may extend this approach by integrating multi-modal data, expanding to other transportation sectors, and exploring temporal sentiment dynamics to further improve service quality and passenger relations.

Gender-Based Analysis of New Year Resolutions on Twitter: A Clustering Approach Using K-means Algorithm

Agung Dharmawan Buchdadi, Ammar Salamh Mujali Al-Rawahna — Mon, 02 Jun 2025 00:00:00 +0700

This study explores gender-based patterns in New Year resolutions shared on Twitter, focusing on differences in resolution content and social media engagement. By analyzing a dataset of 5,011 tweets categorized by user gender, resolution theme, and retweet count, the research aims to uncover how men and women express and share their personal goals in digital spaces. The study employs text preprocessing techniques to clean and normalize tweet texts, followed by Term Frequency-Inverse Document Frequency (TF-IDF) vectorization to convert textual data into numerical features. K-means clustering is then applied to group tweets into five thematic clusters, representing distinct resolution topics. The optimal number of clusters is determined using the Elbow Method, with clustering quality assessed via inertia and Silhouette Score metrics. Results reveal significant gender differences in both resolution categories and engagement patterns. For instance, female users tend to post more about health and fitness or humor-related resolutions, whereas male users show a higher presence in philanthropic and finance-related topics. Although the majority of tweets cluster into broad general resolution themes, specialized clusters reflect focused goal areas like quitting habits and financial improvement. Retweet engagement varies widely, with a skewed distribution where most tweets receive minimal retweets, but select tweets achieve high virality. Gender distributions within clusters were relatively balanced, though some clusters displayed subtle dominance by one gender, highlighting nuanced differences in thematic focus. The study’s findings offer valuable insights into digital self-expression and social sharing behaviors influenced by gender, contributing to the understanding of online social dynamics in the digital society. These insights have practical implications for content personalization, enabling brands and influencers to tailor New Year resolution-related campaigns to gender-specific preferences. Limitations include potential biases in data selection and categorization challenges, suggesting avenues for future research such as longitudinal trend analysis and intersectional demographic studies. Overall, this research advances knowledge on gendered digital communication, highlighting the role of social media in reflecting and shaping personal goal-setting behaviors.

Mining Public Sentiment and Trends in Social Media Discussions on Indonesian Presidential Candidates Using Support Vector Machines

Satrya Fajri Pratama, Arif Mu'amar Wahid — Mon, 02 Jun 2025 00:00:00 +0700

This study investigates public sentiment surrounding the 2024 Indonesian presidential candidates by analyzing Twitter data collected between October 2022 and April 2023. Using Support Vector Machines (SVM) for sentiment classification, the research aims to uncover patterns in early public opinion during the pre-election period. A dataset of 28,782 tweets mentioning three main candidates—Anies Baswedan, Ganjar Pranowo, and Prabowo Subianto—was collected, cleaned, and processed. Exploratory Data Analysis (EDA) revealed that positive sentiment dominated discussions for all candidates, with Ganjar Pranowo receiving the highest proportion of positive tweets. Geographic and user engagement metrics further contextualized the digital discourse, highlighting differences in audience demographics and regional engagement. The text data was transformed using Term Frequency-Inverse Document Frequency (TF-IDF) vectorization, capturing both unigrams and bigrams, and limited to 5,000 features for computational efficiency. The dataset was split into training and testing sets with a 75/25 ratio. A linear kernel SVM was trained with class weighting to address class imbalance, achieving an accuracy of 87.9% on the test set. Precision, recall, and F1-score metrics indicated strong model performance, particularly for the positive sentiment class. Feature importance analysis identified key terms influencing classification, aligning intuitively with sentiment polarity. These findings demonstrate the viability of using SVM and social media data to analyze early political sentiment in emerging democratic contexts. The study contributes to the growing field of digital political analytics by providing insights into voter sentiment dynamics prior to official candidacy announcements. While limited to Twitter and a specific timeframe, the research lays groundwork for multi-platform, longitudinal studies. Future research could explore sentiment evolution throughout the election cycle and deeper analysis of sentiment drivers by region and policy.

Sentiment Analysis of Public Discourse on Pakistan's Political Parties: A Comparative Study Using VADER and TextBlob Algorithms on Twitter Data

Muhamad Irfan, Abdul Sattar, Ahmad Sher, Muhamad Ijaz — Mon, 02 Jun 2025 00:00:00 +0700

This study explores public sentiment toward two of Pakistan's major political parties—Pakistan Peoples Party (PPP) and Pakistan Tehreek-e-Insaf (PTI)—by analyzing Twitter discourse using sentiment analysis techniques. A dataset of 1,184 tweets related to the trending topic "PPP and PTI" was collected and processed to examine how these parties are perceived online. Two lexicon-based sentiment analysis algorithms, VADER and TextBlob, were applied to the tweet content to compute sentiment polarity scores and categorize each tweet as positive, neutral, or negative. Exploratory Data Analysis (EDA) was conducted to assess engagement metrics, tweet length distribution, and user activity patterns. A keyword-based method was used to assign party focus to tweets, enabling comparative sentiment analysis between PPP and PTI. The results indicate that PPP was more frequently mentioned than PTI, comprising over 93% of the classified tweets. Both VADER and TextBlob showed moderate agreement in sentiment classification, with a Pearson correlation coefficient of 0.5761 and a 66.39% match in sentiment labels. Temporal analysis revealed fluctuations in sentiment scores, often corresponding to real-world political events, such as alliance discussions or leadership announcements. Tweets with higher engagement—measured by likes, retweets, replies, and views—tended to exhibit stronger sentiment polarity. Top positive and negative words were also identified to interpret linguistic patterns behind sentiment classification. This study demonstrates the potential of sentiment analysis as a tool for political communication, campaign strategy, and public opinion monitoring. However, limitations such as platform bias and data parsing issues warrant cautious interpretation of the results. Future research may benefit from incorporating multi-platform data and advanced NLP models to enhance reliability and granularity. The findings contribute to the growing field of digital society studies by offering a data-driven lens into political discourse on social media.

Sentiment Analysis of Tweets on Afghan Women’s Rights Using Naive Bayes Classifier: A Data Mining Approach to Understanding Public Discourse

Thosporn Sangsawang, Liu Yang — Mon, 02 Jun 2025 00:00:00 +0700

Social media platforms have become critical arenas for public discourse on global human rights issues, providing real-time insight into public opinion and emotional responses. This study examines Twitter conversations surrounding Afghan women’s rights from April 2023 to January 2024, focusing on the digital reflection of international concern. Using a dataset of 4,845 cleaned English-language tweets, we performed sentiment analysis employing the VADER lexicon for initial sentiment labeling and a Multinomial Naive Bayes classifier trained on TF-IDF features for automated sentiment classification. The results reveal a predominance of negative sentiment (47.4%) compared to positive (38.3%) and neutral (14.3%) sentiments, indicating widespread frustration and alarm regarding the restrictions and violations faced by Afghan women. Exploratory data analysis highlighted temporal trends in tweet volume and engagement, with significant peaks correlating to key political events and policy announcements. The model achieved an overall accuracy of 67.5% in classifying sentiment, with particularly strong performance in detecting negative and positive tweets, while neutral sentiments were more challenging to classify accurately. Feature importance analysis identified critical terms that influenced sentiment classification, revealing a linguistic pattern reflective of advocacy, concern, and hope within the discourse. Temporal analysis of sentiment proportions demonstrated fluctuations aligning with real-world developments, underscoring the dynamic nature of online public opinion. This research contributes to understanding the role of social media in amplifying human rights concerns, especially in politically unstable regions, and demonstrates the utility of sentiment analysis for monitoring global digital activism. The findings offer valuable insights for policymakers, activists, and scholars interested in the intersection of technology, public opinion, and human rights advocacy. Future research is encouraged to incorporate multilingual data, multiple social media platforms, and analyze sentiment shifts in response to international interventions to provide a more comprehensive picture of digital society engagement on this critical issue.