gsh199449 / read-paper

26 stars 3 forks source link

A Unified Model for Extractive and Abstractive Summarization using Inconsistency Loss #48

Open gsh199449 opened 6 years ago

gsh199449 commented 6 years ago

title

A Unified Model for Extractive and Abstractive Summarization using Inconsistency Loss

notes

用抽取式摘要辅助生成式摘要,最终给出生成式摘要的结果,打过了sota。 主要贡献点在于提出了一个inconsistency loss,用来衡量句子维度注意力和词维度注意力的不一致性。然后在loss里面惩罚这种关注度的不一致性。 方法是在解码第t个时刻时,选择这个时刻关注度最大的前k个词,将这k个词的每个词的关注度乘以这个词所在句子的关注度,然后求平均这K个乘积记为。然后再解码的所有时刻的求平均,去在loss中最大化这个值。

bibtex

@inproceedings{Sun2018AUM, title={A Unified Model for Extractive and Abstractive Summarization using Inconsistency Loss}, author={Min Sun and Wan Ting Hsu and Chieh-Kai Lin and Ming-Ying Lee and Kerui Min and Jing Tang}, booktitle={ACL}, year={2018} }

link

https://arxiv.org/abs/1805.06266

publication

ACL 2018 accept

open source

No

affiliated

National Tsing Hua University

paperbot[bot] commented 6 years ago

verified by paper bot