CitiSense: A Web App in Enhancing the Regional Government Agencies Feedback Systems in Pampanga through Facebook Sentiment Analysis with Decision Support Dashboard

Abstract

Regional government agencies in the Philippines typically rely on offline and periodic feedback systems underutilizing vast public opinion expressed on social media. This study addresses this gap by developing “CitiSense”, a web-based decision support dashboard using machine learning and natural language processing to systematically process public sentiment regarding to service delivery. The research gathered code-mixed (Filipino-English) Facebook comments and evaluated four classification models: Random Forest, Multinomial Naïve Bayes, Support Vector Machine, and fine-tuned Multilingual BERT (mBERT). Results demonstrated that mBERT outperformed the three machine learning algorithms, achieving an accuracy of 82% and an F1-score of 0.81. While domain experts and stakeholders found the system highly effective for proactive issue identification, they reported a steep initial learning curve. These findings highlight the potential of sentiment analysis with AI-powered insights to enhance data-driven governance.

Introduction

The paper presents CitiSense, a web-based decision support dashboard designed to transform unstructured social media feedback into actionable intelligence for Philippine regional government agencies. Traditional citizen feedback systems in the Philippines, such as suggestion boxes and periodic surveys, suffer from low participation and delayed insights, despite the country’s high social media usage. Platforms like Facebook generate massive volumes of user-generated content, but analyzing it is challenging due to code-switching ("Taglish"), informal language, sarcasm, and emojis.

Globally, governments have applied automated text analytics, including Multinomial Naive Bayes, for real-time sentiment analysis, highlighting the potential of social media as a proxy for public opinion. Classical machine learning approaches (Naive Bayes, SVM, Random Forest) are limited by manual feature extraction and struggle with nuanced, multilingual text. Transformer-based models, specifically Multilingual BERT (mBERT), handle code-switched and multilingual data effectively, offering superior accuracy in sentiment classification.

Using the CRISP-DM framework, CitiSense collected 14,828 Facebook comments from agencies like DOH, DPWH, DOLE, and DSWD. Data preprocessing involved noise removal, normalization, tokenization, stopword removal, and manual annotation for positive, negative, or neutral sentiment, with sarcasm recoded as negative. Both classical ML models and mBERT were tested, with mBERT fine-tuned for four epochs to handle Filipino-English text.

Overall, CitiSense demonstrates the application of AI-driven, real-time social media analysis to improve governance, institutional accountability, and citizen engagement, aligning with Sustainable Development Goal 16 (Peace, Justice, and Strong Institutions).

Conclusion

The development and evaluation of CitiSense demonstrate that transformer-based architectures, specifically mBERT, provide a robust solution for the technical challenges of analyzing \"Taglish\" sentiment in the Philippine public sector. While traditional manual feedback systems are hindered by delays and low participation, AI-driven dashboards can transform social media into a real-time stream of actionable intelligence. The study revealed three critical insights from the results: 1) Model Superiority: Fine-tuned mBERT achieved a high accuracy of 81.68%, significantly outperforming classical algorithms by better capturing the context of code-switched text and sarcasm. 2) Policy Impact: The system successfully identified service delivery gaps in real-time—such as the bottlenecks in the DSWD’s AICS program—that were absent from traditional monthly reports. 3) The Learnability Paradox: Although the system is highly effective for decision-making, there remains a notable friction point for non-technical administrative staff, who required more intensive onboarding compared to IT personnel. Lastly, the researchers recommended for future work to further enhance the efficacy of CitiSense, future research should focus on developing more intuitive, interactive user guides to bridge the digital literacy gap among government employees. Additionally, expanding the dataset to include audio-to-text feedback from radio programs and community town halls would ensure a more inclusive representation of public sentiment, especially for citizens without consistent internet access. Ultimately, the integration of such AI tools into regional governance serves as a vital step toward achieving Sustainable Development Goal 16, fostering more transparent, responsive, and accountable public institutions.

References

[1] N. A. Alabdulkarim, M. A. Haq and J. Gyani, \"Exploring Sentiment Analysis on Social Media Texts,\" Engineering, Technology & Applied Science Research, 2024. [2] E. Alzahrani and L. Jololian, \"How Different Text- Preprocessing Techniques Using the BERT Model Affect the Gender Profiling of Authors,\" 2021. [3] S. A. Chun and L. F. Luna-Reyes, \"Social media in government,\" Information Polity, vol. 17, no. 3-4, pp. 213-214, 2012. [4] DataReportal, \"Digital 2024: The Philippines,\" Feb. 2024. [5] J. Flor et al., \"Like, Comment, and Share: Analyzing Public Sentiments of Government Policies in Social Media,\" n.d.. [6] L. L. Maceda, A. A. Satuito and M. B. Abisado, \"Sentiment Analysis of Code-mixed Social Media Data on Philippine UAQTE using Fine-tuned mBERT Model,\" IJACSA, vol. 14, no. 7, 2023. [7] K. Karoo and V. Chitte, \"Ethical Considerations in Sentiment Analysis,\" IRJMETS, 2023. [8] D. J. S. Oliveira et al., \"The application of the sentiment analysis technique in social media,\" Revista de Administração Pública, vol. 53, no. 1, pp. 235–251, 2019. [9] C. A. Pacol and T. D. Palaoag, \"Enhancing Sentiment Analysis of Textual Feedback,\" Euro. J. Eng. Sci. Tech., vol. 4, no. 1, pp. 27-34, 2021. [10] N. Sande, I. O. Adeniyi and A. Akinkunmi, \"Social Media Sentiment Analysis: A Comprehensive Analysis,\" 2024. [11] D. F. Putra and Y. Sibaroni, \"Sentiment Analysis for the 2024 Presidential Election Using BERT CNN,\" 2024. [12] J. M. Umali, J. P. P. Miranda and A. L. Ferrer, \"Sentiment analysis: A case study among the selected government agencies in the Philippines,\" IJATCSE, vol. 9, no. 3, pp. 3492–3497, 2020. [13] W. B. Zulfikar, A. R. Atmadja and S. F. Pratama, \"Sentiment Analysis on Social Media Against Public Policy Using Multinomial Naive Bayes,\" Scientific Journal of Informatics, vol. 10, no. 1, pp. 25–34, 2023. [14] M. Sethi et al., \"Sentiment identification in COVID-19 specific tweets,\" ICIEM, pp. 363–367, 2020

Copyright

Copyright © 2026 Mark Gio G. Alcuizar, Haidee Adreanne F. Duarte, Angeline L. Ruin, John Ryan P. Trinidad, Joey S. Aviles. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.