Disaster Tweet Classification

NLP model using BERT to classify disaster-related tweets

Overview

Built a natural language processing model to classify tweets into disaster and non-disaster categories, enabling rapid identification of emergency situations from social media.

Technical Details

  • Technologies: Python, TensorFlow, Hugging Face Transformers
  • Model: BERT (Bidirectional Encoder Representations from Transformers)
  • F1 Score: 80.90%

Approach

  1. Text Preprocessing: Cleaned tweets, handled URLs, mentions, and hashtags
  2. Tokenization: Used BERT tokenizer for subword tokenization
  3. Fine-tuning: Fine-tuned pre-trained BERT on the disaster classification task
  4. Evaluation: Used F1 score as primary metric due to class imbalance

Results

The BERT model achieved an F1 score of 80.90%, demonstrating strong performance in distinguishing genuine disaster reports from figurative language.

Applications

This model can be used by:

  • Emergency response teams for real-time disaster monitoring
  • News organizations for breaking news detection
  • Government agencies for crisis management