Social Media Analysis is an AI system that automatically classifies content from X (formerly Twitter) to support the Water Authority’s leadership. It replaces a 4-hour daily manual process with automated sentiment, location, and problem classification. By feeding daily insights into the Observatory Dashboard, it enables higher management to understand citizen feedback and service reports instantly, significantly improving digital transparency and public engagement.
Initiative overview
The Social Media Analysis initiative was developed to bridge the gap between massive public discourse on social media and actionable institutional intelligence. Previously, monitoring citizen feedback on X was a labor-intensive manual process requiring roughly four hours daily. This initiative leverages AI and machine learning to automate this workflow, processing data through Google Cloud BigQuery and visualising results via Looker Enterprise. By doing so, it resolves the issue of "information overload" and ensures that critical beneficiary reports are never missed.
The primary objective is to transform raw social data into structured business insights through several specialised layers: Spam Guard: Uses NLP to filter out advertisements, water filter promotions, and redundant bot replies. Multi-Layer Classification: The system assigns sentiment (Positive, Neutral, Negative), extracts geographic data (City and Neighbourhood), and maps posts to approximately seven business-defined problem categories. Summary Reporting: The data points are synthesised and integrated into the "Observatory Dashboard," providing a single source of truth for leadership and higher management, highlighting emerging trends and public perception.
The initiative is currently fully operational and success is maintained through a combination of automated infrastructure monitoring on Google Cloud Run and continuous human-in-the-loop validation to ensure the model remains aligned with local context and evolving dialect. Future evolution of the tool is focused on horizontal and functional scaling. The roadmap includes expanding analysis to platforms like Facebook, LinkedIn and Instagram, as well as incorporating video and Reels content analysis. Additionally, we aim to integrate the system directly with CRM platforms to automatically trigger service tickets from social media complaints, further closing the loop between citizen feedback and operational response.
Other relevant details
Results, outcomes and impacts: The initiative has achieved a 95% average classification accuracy (ranging 83–98%), reducing a 4-hour manual task to a 10–20 second automated process per 100 posts. We maintain 99.3% system uptime and an error rate below 2.5%. Impact is measured via an automated GCP monitoring framework (Cloud Run/Pub/Sub) and human audits by the Observatory team, showing a 92% feedback resolution rate. Future impacts include expansion to all major social platforms and integrated CRM ticketing. These tangible results ensure a high-speed, data-consistent environment for management, with data consistency levels exceeding 97% across all processed batches.
Challenges and lessons learned: A significant challenge was geographic data bias, as social media activity naturally skews toward major cities. We addressed this through keyword expansion and location-based filtering to ensure smaller regions were represented fairly. Filtering "noise" required a sophisticated Spam Guard to distinguish between meaningful feedback and automated bot replies. Key Lessons: Trust through Transparency: Accuracy is verified using a strict formula by calculating the percentage of correct predictions among total predictions. Data is King: The project’s success was primarily driven by a high-quality training dataset and constant model testing. Success requires strong infrastructure (GCP), high-quality data cleaning, and adherence to SDAIA’s fairness principles to ensure classification neutrality.
