headlines

Why we included this project

Headline generation is a fiddly problem, and this repo is one of the clearer worked examples of doing it with Keras. It reproduces the approach from the paper 'Generating News Headlines with Recurrent Neural Networks', training an attention model on hundreds of thousands of paired article openings and original headlines. A nice touch that goes beyond the paper is copying out-of-vocabulary words from the article text directly into the generated headline, which helps when proper nouns carry the story. The Jupyter notebooks walk you through building a GloVe embedding vocabulary, training, and then inspecting attention weights to see which description words drove each title. It is research example code rather than a production service, so you will need to bring your own data and adapt it, but it is a readable starting point for anyone experimenting with summary-style generation.