Enhancing Label Correlation Feedback in Multi-Label Text Classification via Multi-Task Learning

This repository contains the code for the ACL 2021 paper

"Enhancing Label Correlation Feedback in Multi-Label Text Classification via Multi-Task Learning".

If you use LACO in your work, please cite it as follows:

@article{zhang2021enhancing,
  title={Enhancing Label Correlation Feedback in Multi-Label Text Classification via Multi-Task Learning},
  author={Zhang, Ximing and Zhang, Qian-Wen and Yan, Zhao and Liu, Ruifang and Cao, Yunbo},
  journal={arXiv preprint arXiv:2106.03103},
  year={2021}
}

Settings

Environment Requirements

• python 3.6+

• Tensorflow 1.12.0+

Environmental preparation

• You can change the experimental settings in LACO/common/global_config.py

• The initial content under directory LACO/ie/src/bert is primarily from Google bert. Citation information is recorded in the corresponding file. You can download and unzip it at LACO/pretrained_model/ .

Datasets

• AAPD
• RCV1-V2

Data Preparation

The sample data are in the directory LACO/log/re_model/input. Note that the "text" field stores the text content, the "spo_list" field stores the relevant labels in "predicate", and the other fields can be ignored.

How To Run

• Train_mltc_with_plcp: python ie/train_main_plcp.py
• Test_mltc_with_plcp: python ie/test_main_plcp.py
• Train_mltc_with_clcp: python ie/train_main_clcp.py
• Test_mltc_with_clcp: python ie/test_main_clcp.py

Results

The best model of +PLCP of AAPD dataset and and RCV1V2 dataset can be found at https://share.weiyun.com/5EXHqEPE (password: 8yrgji) for your reference.

© Copyright

Ximing Zhang (ximingzhang@bupt.edu.cn),
Qian-Wen Zhang (cowenzhang@tencent.com),
Zhao Yan (zhaoyan@tencent.com),
Ruifang Liu (lrf@bupt.edu.cn),
Yunbo Cao (yunbocao@tencent.com),
Tencent Cloud Xiaowei, Beijing, China  && Beijing University of Posts and Telecommunications, Beijing, China  

This code package can be used freely for academic, non-profit purposes. For other usage, please contact us for further information (Ximing Zhang: ximingzhang@bupt.edu.cn).