udibr/headlines

Implements a sequence-to-sequence RNN architecture with attention mechanism for abstractive headline generation, based on the 2015 "Generating News Headlines with Recurrent Neural Networks" paper. Uses Keras for model training and incorporates GloVe embeddings for word representations, with a notable extension enabling out-of-vocabulary word copying directly from source articles. Provides end-to-end pipeline including vocabulary building, model training via Jupyter notebooks, and inference with visualization of attention weights.

524 stars. No commits in the last 6 months.