Civic Complaint Classification System Using DistilBERT

A multilingual AI-powered civic complaint management system that automatically categorizes citizen complaints in English and Nepali using fine-tuned DistilBERT models.

POSTER OF CIVIC COMPLAINT CLASSIFICATION SYSTEM USING DISTILBERT

Overview

Manual complaint routing in public services leads to delays and misrouting. This system addresses that by automatically classifying citizen-submitted text complaints into predefined categories, routing them to the relevant authorities in real time.

Complaint Categories:

• Road
• Water
• Electricity
• Sanitation / Garbage

Key Highlights:

• Supports both English and Nepali language inputs
• English model achieves ~99.8% accuracy on validation data
• Nepali model delivers strong performance despite limited training data
• Integrated with a Django REST API backend and Flutter mobile app

---

Project Structure

.
├── backend/                        # Django REST API backend
│   └── mysite/
│       ├── accounts/               # User authentication
│       ├── authority/              # Authority/admin management
│       ├── citizen/                # Citizen profile & complaint submission
│       ├── complaint/              # Complaint handling & ML inference
│       └── ml_models/              # Trained DistilBERT model directories
│           ├── civicconnect/       # English complaint classifier
│           └── nepali/             # Nepali complaint classifier
│
├── civiccomplaints/                # Flutter mobile application
│   └── lib/
│       ├── screens/                # App screens (login, home, submit, etc.)
│       ├── services/               # API service calls
│       ├── widgets/                # Reusable UI components
│       └── theme/                  # App theme
│
├── project/                        # ML training & evaluation scripts
│   ├── civicconnect/               # English model pipeline
│   │   ├── train_distilbert.py
│   │   ├── predict_distilbert.py
│   │   └── evaluate_hard_test.py
│   └── nepali/                     # Nepali model pipeline
│       ├── train_distilbert.py
│       ├── predict_distilbert.py
│       └── generate_complaints.py
│
└── assets/                         # Repository assets (poster, images)

---

Tech Stack

Layer	Technology
Mobile App	Flutter 3.x (Dart)
Backend API	Django 6.0 (Python)
ML Models	DistilBERT (HuggingFace Transformers)
Database	SQLite (dev)
Tunneling (dev)	ngrok

---

Getting Started

Prerequisites

• Python 3.10+
• Flutter SDK 3.x
• pip

Backend Setup

cd backend/mysite
pip install -r requirements.txt    # install Django & ML dependencies
python manage.py migrate
python manage.py runserver

Note: Model weight files (.safetensors) are not included in this repository due to size limits. Place the trained model files at:

• backend/mysite/ml_models/civicconnect/best_model/model.safetensors
• backend/mysite/ml_models/nepali/best_model/model.safetensors

Flutter App Setup

cd civiccomplaints
flutter pub get
flutter run

Update the API base URL in civiccomplaints/lib/services/ to match your backend address (or ngrok URL for local testing).

Training the Models

cd project/civicconnect      # or project/nepali
pip install -r requirements.txt
python train_distilbert.py

---

Model Performance

Model	Accuracy	F1 (Weighted)
English (CivicConnect)	~99.8%	~0.998
Nepali	High	Reliable across 4 classes

Evaluation includes confusion matrices, precision/recall/F1 per class, and training/validation loss curves — see project/civicconnect/results/ and project/nepali/results/.

---

License

This project is for academic research purposes. See individual component directories for any additional license information.