attention-networks-for-classification

Why we included this project

This is a PyTorch implementation of the Hierarchical Attention Network from the NAACL 2016 paper, built around the idea that documents have a natural structure: words combine into sentences, and sentences into documents. Instead of letting the model discover that structure on its own, the code feeds it in explicitly through word-level and sentence-level attention over a bidirectional GRU. The included notebook trains the model on a movie review dataset, so you get a complete, runnable example of the pipeline rather than isolated snippets. It's a good starting point for researchers or students who want to see how attention-based document classification actually works under the hood. Just keep in mind the repo hasn't seen updates since early 2020, so treat it as a teaching tool and a base for your own experiments, not something to wire into production.