BertSum

Why we included this project

BertSum is the code that accompanies a frequently cited paper on extractive summarization, the kind where a model picks the strongest existing sentences from a document instead of writing new ones. It shows how to put a sequence-labeling layer on top of a pretrained BERT encoder, score each sentence, and assemble the top-scoring ones into a summary. Anyone reproducing or extending that line of work will appreciate that the repo ships the full CNN/DailyMail preprocessing pipeline and ROUGE evaluation scripts. The trade-off is age: it targets Python 3.6 and the pytorch_pretrained_bert library, so expect to port pieces before it runs on modern PyTorch. For learning how to fine-tune BERT for sentence selection, though, the code is still a useful place to start.