@inproceedings{Wang2018ART,
title={A Reinforced Topic-Aware Convolutional Sequence-to-Sequence Model for Abstractive Text Summarization},
author={Li Wang and Junlin Yao and Yunzhe Tao and Li Zhong and Wei Liu and Qiang Du},
booktitle={IJCAI},
year={2018},
pages={4453--4460},
}
Wanghn95
commented
5 years ago
title
A Reinforced Topic-Aware Convolutional Sequence-to-Sequence Model for Abstractive Text Summarization
notes
做生成式短文本摘要，在DUC， Gigaword和LCSTS打到stoa。方法是用ConvS2S作为基础，在attention的时候不仅attend到原文的词，还attend到topic的词。其中topic是预先用LDA生成的词。在训练的时候使用RL优化ROUGE，详细说了为啥要用RL，比MLE好在哪。MLE一是训练的时候只让模型暴露在真实数据中，而没有把自己的输出给下一步，而测试的时候是用自己的输出。二是优化MLE只有输出的和真实值一模一样的时候才行，及时语义一致都不行。而有多个reference的时候用ROUGE做reward却可以。
bibtex
@inproceedings{Wang2018ART, title={A Reinforced Topic-Aware Convolutional Sequence-to-Sequence Model for Abstractive Text Summarization}, author={Li Wang and Junlin Yao and Yunzhe Tao and Li Zhong and Wei Liu and Qiang Du}, booktitle={IJCAI}, year={2018}, pages={4453--4460}, }
link
https://www.ijcai.org/proceedings/2018/0619.pdf
publication
IJCAI 2018 long accepted
open source
No
affiliated
Tencent SNG, AI Lab