An NLP and deep learning project that classifies news articles into categories — exploring LSTM, BERT, and transformer-based architectures to find the best approach.
Automatically categorizing news content at scale is a core NLP challenge — one that requires understanding context, tone, and subject matter, not just keywords.
Trained and compared multiple architectures — starting with LSTM for sequence modeling, then fine-tuning BERT to leverage pre-trained contextual understanding.
A working multi-class text classifier with [your best accuracy here]% accuracy on the HuffPost dataset, with a clear analysis of the tradeoffs between model types.
The HuffPost dataset is a well-known benchmark for NLP classification tasks — it contains news headlines and short descriptions across a range of categories, making it a strong test for models that need to understand topic and context.
I approached this as an architecture comparison study as much as a classification task. Starting with a baseline TF-IDF + logistic regression model, I worked up to bidirectional LSTM for sequential modeling, then fine-tuned a BERT model for the best contextual understanding. The goal was to understand not just which model performs best, but why — and what that means for real-world NLP decisions.
Add your specific findings and observations here — what surprised you, what didn't work as expected, what you'd do differently.
Baseline (TF-IDF + Logistic Regression) — Accuracy: [add yours]. Fast, interpretable, surprisingly competitive.
Bidirectional LSTM — Accuracy: [add yours]. Better at capturing sequence context within headlines.
BERT (fine-tuned) — Accuracy: [add yours]. Best overall, especially on ambiguous categories.