UPDATE: This paper is part of the our EMNLP2020 findings: paper code.
- Model bias: CopyMTL suffers from the exposure bias problem, which can be solved by our Seq2UMTree.
- Data bias: NYT dataset is overfitted by SoTA models. This is because 90% test triplets reoccured in the training data.
- We release OpenJERE toolkit, including multiple baselines and datasets. CopyMTL can be found here!
CopyMTL: Copy Mechanism for Joint Extraction of Entities and Relations with Multi-Task Learning
Paper accepted by AAAI-2020
This is a followup paper of "Extracting Relational Facts by an End-to-End Neural Model with Copy Mechanism" ACL2018 CopyRE
Environment
python3
pytorch 0.4.0 -- 1.3.1
Modify the Data path
This repo initially contain webnlg, you can run the code directly. NYT dataset need to be downloaded and to be placed in proper path. see const.py.
The pre-processed data is avaliable in:
WebNLG dataset: https://drive.google.com/open?id=1zISxYa-8ROe2Zv8iRc82jY9QsQrfY1Vj
NYT dataset: https://drive.google.com/open?id=10f24s9gM7NdyO3z5OqQxJgYud4NnCJg3
Run
python main.py --gpu 0 --mode train --cell lstm --decoder_type one
python main.py --gpu 0 --mode test --cell lstm --decoder_type one