fduyjx/rsman

Source code for NAACL 2022 paper: Relation-Specific Attentions over Entity Mentions for Enhanced Document-Level Relation Extraction

RSMAN

Source code for NAACL 2022 paper: Relation-Specific Attentions over Entity Mentions for Enhanced Document-Level Relation Extraction

Introduction

This is the implementation of the RSMAN (Relation-Specific Mention Attention Network) with SSAN as the backbone model. RSMAN is easy to be plugged into other different backbone models to enhance them and here we take SSAN for example. Part of the code is borrowed from https://github.com/BenfengXu/SSAN, and we really appreciate it.

Requirements

• python==3.6
• pytorch==1.4.0
• transformers==2.7.0

Dataset

• DocRED
• DWIE
• Note that you should process DWIE to fit the same format as DocRED. Put the dataset into the directory ./data.

Train

Download pre-trained language models into the directory ./pretrained_lm and run:

python run.py

The evaluation on dev set will be run during training at each logging step, and the trained model corresponding to the best dev result will be saved into the directory ./checkpoints.

Test

To get the result on test set, run:

python run.py --do_train False --do_predict

Then a test result file in the official evaluation format will be saved as ./checkpoints/result.json. Compress and submit it to CodaLab to get the final test score.

We also provide our trained model and test result file, you can download them from here.

rsman's Issues

where can i find your paper?

DWIE数据集处理问题

您好，您的这篇论文写的很好，看了你的文章对我很有启发。
请问是否可以公开您将DWIE数据集处理为DocRED 格式的代码？这将对我会很有帮助。
方便的话可以发送到我的邮箱，[email protected]
期待你的回复，不胜感激！

DWIE训练结果求助

您好，之前看到您在LogiRE论文的Issue提到，BiLSTM的F1只能到40左右，但您论文中还是引用了LogiRE的数据，请问后来BiLSTM能得到接近50%的结果了么？