Disaster Tweets — NLP Architecture Benchmarks.

Overview

A controlled NLP benchmarking study across 7 architectures — TF-IDF + Naive Bayes, Dense, LSTM, GRU, BiLSTM, Conv1D, and Universal Sentence Encoder transfer learning — on ~7,600 disaster tweets with a consistent 90/10 split. USE transfer learning achieved the best results (81.5% accuracy, 0.81 F1), demonstrating that pre-trained encoders outperform from-scratch models on small datasets.

Highlights

• +7 architectures under identical splits: classical ML, RNN family (LSTM/GRU/BiLSTM), Conv1D, and USE transfer learning
• +Best model: Universal Sentence Encoder transfer learning — 81.5% accuracy, 0.81 F1
• +Clear evidence that pre-trained encoders beat from-scratch models on small text datasets

Stack

Python · TensorFlow · TensorFlow Hub · scikit-learn · Pandas