Official source code for Towards Interpretable Natural Language Understanding with Explanations as Latent Variables.

Directory ELV_re and ELV_sa is source code for Relation Extraction task and Sentimental Analysis task in supervised setting(Use all classification labels). There are some tiny difference in training details and pre-processing. Directory EST_re and EST_sa are codes in semi-supervised setting(Use part of classificatioin labels).

Directory data is data used in our experiment. TACRED is not released because of its copyright. It can be bought and downloaded at LDC TACRED webpage.

To replicate the result of experiment, run the bash script in each directory. For example, to replicate the ELV result on Semeval dataset, just use

bash train_semeval.sh

We implement Bert classifier based on Huggingface transformers and Unilm generator based on Microsoft Unilm.

@article{zhou2020towards,
  title={Towards Interpretable Natural Language Understanding with Explanations as Latent Variables},
  author={Zhou, Wangchunshu and Hu, Jinyi and Zhang, Hanlin and Liang, Xiaodan and Sun, Maosong and Xiong, Chenyan and Tang, Jian},
  journal={arXiv preprint arXiv:2011.05268},
  year={2020}
}