Implements the model described in the following paper End-to-End Abstractive Summarization for Meetings.
@article{zhu2020end,
title={End-to-End Abstractive Summarization for Meetings},
author={Zhu, Chenguang and Xu, Ruochen and Zeng, Michael and Huang, Xuedong},
journal={arXiv preprint arXiv:2004.02016},
year={2020}
}
1. I used CNN in Transformer-PositionwiseFeedForward.
2. Role vectors, pos tags, and named entity tags are not used.
(When using a role vector, performance was lower than not utilizing role vector. So please do not hesitate to advise me about this.)
python main.py --mode train --save_path path_to_save_the_model
python main.py --mode eval --model_path trained_model_path --gen_max_length 500
Epoch | Rouge-1 | Rouge-2 | Rouge-L |
---|---|---|---|
30 | 0.4762 | 0.1862 | 0.1767 |
40 | 0.4796 | 0.1935 | 0.1858 |