PAMM-HiA-T5 consists of the Hierarchy-Aware T5 and the Path-Adaptive Mask Mechanism. The project consists of following parts:
- data: Data dir for the preprocessed RCV1, NYT, WOS datasets (because of the datasets' size exceeds the available max limit set by ARR, we only upload a representative subset of them). The original datasets could refer to RCV1-V2, NYT and WOS.
- pretrain_model: Download the relevant files of the pre-training T5 model including pytorch_model.bin, config.json, tokenizer.json, spiece.model, etc. from T5-base, and then put them in the project path: PAMM-HiA-T5/pretrain_model/t5-base.
- utils.py: The data processing and data loader of PAMM-HiA-T5.
- dmask.model_t5_4_classification & train_dmask.py: The main model of PAMM-HiA-T5 and its training script.
- train.py: The main model of HiA-T5 and its training script.
- test.py: The test script of PAMM-HiA-T5 or HiA-T5.
- python 3.7.9
- pytorch 1.7.0
- transformers 2.9.0
The hyperparameters of PAMM-HiA-T5 are configured in the args_dict
of train_dmask.py
. You can change all hyperparameters and run train_dmask.py
to train PAMM-HiA-T5 on different settings. To test the model, you can change the ckpt_path
, dataset
, and badcase_path
in test.py
and then run test.py
.
https://aclanthology.org/2022.coling-1.95/
Wei Huang, Chen Liu, Bo Xiao, Yihua Zhao, Zhaoming Pan, Zhimin Zhang, Xinyun Yang, and Guiquan Liu. 2022. Exploring Label Hierarchy in a Generative Way for Hierarchical Text Classification. In Proceedings of the 29th International Conference on Computational Linguistics, pages 1116–1127, Gyeongju, Republic of Korea. International Committee on Computational Linguistics.