repozhang / MCRF

Improving Multi-label Malevolence Detection in Dialogues through Multi-faceted Label Correlation Enhancement

Dependencies

• csv
• tqdm
• numpy
• pickle
• scikit-learn
• PyTorch1.5+
• matplotlib
• pandas
• transformers=4.5.1

How to use the code

data

data preprocessing

We use MDMDdata which is uploaded to openview.net

1. Put the dataset file in the ./pybert/dataset/data_transform/data_input/multi-label-n

2. Use ./pybert/dataset/data_transform/data_preprocessiong_n.py to split the data, the splited data will saved in ./pybert/dataset/data_transform/data_set/

3. cd pybert/dataset/data_transform

4. Run python data_preprocessiong_n to preprocess the data.

bert

Bert model is part of the bert-MCRF model and is a baseline model.

you need to download pretrained bert model firstly.

BERT: bert-base-uncased

1. Download the Bert pretrained model from here

2. Download the Bert config file from here

3. Download the Bert vocab file from here

4. Rename:

   • bert-base-uncased-pytorch_model.bin to pytorch_model.bin
   • bert-base-uncased-config.json to config.json
   • bert-base-uncased-vocab.txt to bert_vocab.txt

5. Place model ,config and vocab file into the /pybert/pretrain/bert/base-uncased directory.

6. pip install pytorch-transformers from github.

7. We use MDMDdata, which is uploaded to openview.net.

8. Modify configuration information in pybert/configs/basic_config.py(the path of data,...).

9. Run python run_bert.py --do_data to transfer the splited data to formulated data which will be save in ./pybert/dataset/data_transform/data_set/

10. Run python run_bert.py --do_train --save_best --do_lower_case to fine tuning bert model.

11. Run run_bert.py --do_test --do_lower_case to predict new data.

12. After training the bert model, we move the the trained bert model from ./pybert/output/checkpoints/bert to pybert/model/prev_trained_model/bert-base and continue the save the Bert-MCRF model.

bert-MCRF

Run cd pybert/model to prepare to train the bert-MCRF model.

1. Set TASK_NAME="mcrf" in ./pybert/model/models/scripts/run_bert_MCRF.sh
2. Set --do_train
3. run bash scripts/run_bert_MCRF.sh with 2 GPUs to train the bert-MCRF model.
4. Set --predict
5. run bash scripts/run_bert_MCRF.sh with 2 GPUs to train the bert-MCRF model.

bert-CRF

Bert-CRF is a baseline model.

Run cd pybert/model to prepare to train the bert-CRF model.

1. Set TASK_NAME="crf" in ./pybert/model/models/scripts/run_bert_CRF.sh
2. Set --do_train
3. run bash scripts/run_bert_CRF.sh with 2 GPUs to train the bert-CRF model.
4. Set --predict
5. run bash scripts/run_bert_CRF.sh with 2 GPUs to train the bert-CRF model.

Ablation study

w/o LCC

1. Set TASK_NAME="mcrfwolcc" in ./pybert/model/scripts/run_bert_MCRF.sh
2. In ./pybert/model/models/bert_MCRF.py change the from .layers.CRFKL import CRF with from .layers.CRFKLwoLCC import CRF
3. run bash scripts/run_bert_MCRF.sh with 2 GPUs to train the bert-MCRF model.
4. Set --predict
5. run bash scripts/run_bert_MCRF.sh with 2 GPUs to train the bert-MCRF model.

w/o LCT

1. Set TASK_NAME="mcrfwolct" in ./pybert/model/scripts/run_bert_MCRF.sh
2. In ./pybert/model/models/bert_MCRF.py delete position_ids=position_ids in line 78.
3. Use line99 and delete line101 and line 102.
4. run bash scripts/run_bert_MCRF.sh with 2 GPUs to train the bert-MCRF model.
5. Set --predict
6. run bash scripts/run_bert_MCRF.sh with 2 GPUs to train the bert-MCRF model.

w/o LLCT

1. Set TASK_NAME="mcrfwolct" in `./pybert/model/scripts/run_bert_MCRF.sh
2. Use line99 and delete line101 and line 102.
3. run bash scripts/run_bert_MCRF.sh with 2 GPUs to train the bert-MCRF model.
4. Set --predict
5. run bash scripts/run_bert_MCRF.sh with 2 GPUs to train the bert-MCRF model.

w/o PLCT

1. Set TASK_NAME="mcrfwolct" in ./pybert/model/scripts/run_bert_MCRF.sh
2. In ./pybert/model/models/bert_MCRF.py delete position_ids=position_ids in line 78.
3. run bash scripts/run_bert_MCRF.sh with 2 GPUs to train the bert-MCRF model.
4. Set --predict
5. run bash scripts/run_bert_MCRF.sh with 2 GPUs to train the bert-MCRF model.

Please refer to our paper:

@inproceedings{zhang2022improving, title={Improving Multi-label Malevolence Detection in Dialogues through Multi-faceted Label Correlation Enhancement}, author={Zhang, Yangjun and Ren, Pengjie and Deng, Wentao and Chen, Zhumin and Rijke, Maarten}, booktitle={Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)}, pages={3543--3555}, year={2022} }