Extractive Document Summarizers - SideNet and Refresh

Description

This is a demo of SideNet and Refresh, our two neural extractive document summarization systems.

SideNet builds the summary by extracting relevant sentences from the document by attending side information such as title and captions in the document. SideNet’s TensorFlow source code can be found on github.

Refresh frames extractive summarization as a sentence ranking task and uses a novel training algorithm which globally optimizes the ROUGE evaluation metric through a reinforcement learning objective.

Both systems were developed by Shashi Narayan (in collaboration with Nikos Papasarantopoulos, Mirella Lapata and Shay Cohen).

We thank our funders: EU H2020 (grant agreement: 688139, SUMMA) and Huawei Technologies.

Papers

To read more about SideNet, see the paper that can be downloaded here.

@article{sidenet-2017,
  author={Shashi Narayan and Nikos Papasarantopoulos and Mirella Lapata and Shay B. Cohen},
  title={Neural Extractive Summarization with Side Information},
  journal={CoRR},
  volume={abs/1704.04530},
  year={2017},
  ee={http://arxiv.org/abs/1704.04530},
}

To read more about Refresh, see the paper that can be downloaded here.

@inproceedings{narayan-18,
  author={Shashi Narayan and Shay B. Cohen and Mirella Lapata},
  title={Ranking Sentences for Extractive Summarization with Reinforcement Learning},
  booktitle = {Proceedings of NAACL},
  year = 2018
 }

Demo

Newstype:

URL of article:




Text:



LEAD Summary Baseline (first three sentences in the article)

Gold Summary

Sidenet Summary

Refresh Summary

LOG

LOG