vietnh1009/Hierarchical-attention-networks-pytorch

Implements two-level attention mechanisms at word and sentence levels to capture document structure, with GloVe word embeddings (50-300d) initialized in the embedding layer rather than default random initialization. Built on PyTorch with early stopping regularization and TensorBoard integration for training visualization. Includes a web demo interface and pre-trained models evaluated across eight datasets (AG News, DBPedia, Yelp, Amazon, Yahoo Answers) with configurable batch size, learning rate, and embedding dimensions.

406 stars. No commits in the last 6 months.