AI4SDG-BERT is a fine-tuned BERT model for detecting disaster-related content in tweets. It supports real-time monitoring of social media data to identify and classify potential urban disasters. This model contributes to Sustainable Development Goal 11 (Sustainable Cities and Communities) by facilitating early warning and emergency response systems.

Intended Use

This model is intended for:

Disaster detection from Twitter or other social platforms
Urban emergency alert systems
Smart city infrastructure monitoring
Use by NGOs, civil defense units, or emergency data pipelines

Model Architecture

Base model: bert-base-uncased
Fine-tuned on: Disaster Tweet Classification Dataset (binary classification: disaster vs. not disaster)
Framework: Hugging Face Transformers + PyTorch
Post-BERT Classifier: Random Forest trained on BERT [CLS] embeddings

Training Dataset

Class distribution: 43% disaster, 57% non-disaster
Preprocessing: Lowercase, punctuation/emoji removal, URL stripping

Evaluation

Evaluated on held-out test set from the dataset:

Accuracy: 92%
F1-score: 91%
AUC: 94%

Limitations

English-only model
May not generalize well to sarcasm, spam, or very informal text
Lacks direct geolocation unless paired with NER/metadata
Relies on Twitter API access for live inference

Ethical Considerations

May reflect bias in who reports disasters and how
Not intended as a sole source for life-critical decision making
Requires moderation to avoid misinformation risks

Example Usage

python

from transformers import pipeline
classifier = pipeline("text-classification", model="elam2909/bert-disaster-classifier")
classifier("Huge earthquake just hit southern Turkey. Buildings are collapsing!")