Wei Li

引入预训练有两种方式：（1）每个段落单独用RoBERTa编码；（2）所有段落按论文中所示拼接成序列，长度可以放缩，不用截断512，position embedding也需要放缩，可参考https://github.com/nlpyang/PreSumm/blob/master/src/models/model_builder.py#L200 论文中是采用第(2)种方式的结果，第一种方式也可以，结果比较接近。

中文模型准备中，也会开源

Thanks for your attention. Could you tell me what's the difference between your dataset and the one I provided? I will check it.

The test set has two parts MultiNews.30.test.0.json and MultiNews.30.test.1.json. Maybe the order of the data instances are different. So I think your script could not reflect the difference. Could you...

The discourse graph is constructed based on http://knowitall.cs.washington.edu/gflow/

It shows that your PaddlePaddle env has problems. Please check it. You can install PaddlePaddle following https://www.paddlepaddle.org.cn/documentation/docs/zh/install/index_cn.html

Yes, of course. You can train GraphSum on your own dataset.

这里graph_attn_bias就是1-G[i][j]，这块在reader里面转换了https://github.com/PaddlePaddle/Research/blob/f745a7a5668d68ef7fe5183b67cb1fb3c8eff25c/NLP/ACL2020-GraphSum/src/networks/graphsum/graphsum_reader.py#L368 

Sorry for the delay reply. Because the paper was during the anonymity period of ACL2021. The paper has been updated a lot and has been accepted by the main conference...

The repo for UNIMO has been moved to https://github.com/PaddlePaddle/Research/tree/master/NLP/UNIMO