sadeem-ai/citizen-feedback-classifier
Citizen Feedback Classifier
Macro F1 (topic)
0.88
F1 (sentiment)
0.86
Accuracy (urgency)
0.91
Macro F1 (Gulf dialect)
0.84
About this model
Multi-task Arabic transformer that classifies citizen feedback by topic, sentiment, and urgency across 14 government service categories. Winner of the 2025 Arabic Dialect Sentiment Challenge, with strong performance on Gulf-dialect and code-switched submissions.
Intended use
Routing and triage of feedback from government portals, contact centres, and the Oun app; aggregated sentiment reporting for service-quality dashboards.
Training data lineage
Fine-tuned on the citizen-feedback-corpus of 480,000 labelled submissions collected across national service channels between 2023 and 2025.
Limitations & bias notes
Sarcasm and mixed-sentiment submissions remain the largest error source, particularly in Gulf-dialect posts. Categories introduced after training (e.g. new e-services) fall back to a generic topic label until the next fine-tune.
Evaluation metrics
| Macro F1 (topic) | 0.88 |
| F1 (sentiment) | 0.86 |
| Accuracy (urgency) | 0.91 |
| Macro F1 (Gulf dialect) | 0.84 |
Try it
Live sandboxFeedback
Owning entity
- Updated
- 2026-07-12
- Latest version
- 2.3.0
- License
- Revenue Share
- Access
- Open
Trained on
citizen-feedback-corpus →