Disaster Tweet Classification

• 

  This is the screenshot of the validation for the model which achieved an F1 score of = 0.8955

Inspiration

Real disasters generate thousands of social media posts per minute. First responders need automated tools to filter actionable information from noise. This motivated building a high-accuracy disaster tweet classifier.

What I Built

A binary text classifier using BERTweet (vinai/bertweet-base) — pretrained on 850 million tweets — fine-tuned to classify disaster tweets as Informative or Not Informative.

How I Built It

• Combined Tweet Text + Information Source + Information Type as input features
• Fine-tuned BERTweet for 3 epochs on 25,933 labeled tweets
• 90/10 train-validation split with seed=42 for reproducibility
• Platform: Google Colab T4 GPU

Results

Informative → Precision: 0.92 | Recall: 0.93 | F1: 0.92

Not Informative → Precision: 0.87 | Recall: 0.86 | F1: 0.87

✅ Macro F1-Score: 0.8955

Challenges Faced

• Choosing domain-specific BERTweet over generic BERT for better tweet understanding
• Combining multiple text columns for richer feature signal

What I Learned

Domain-specific pretraining significantly outperforms generic models for social media text classification.

Built With

• bertweet
• google-colab
• huggingface-transformers
• numpy
• pandas
• python
• pytorch
• scikit-learn
• t4-gpu