laddie132/lw-pt

Dataset and code for "Label-Wise Document Pre-Training for Multi-Label Text Classification" (NLPCC 2020)

Label-Wise Pre-Training (LW-PT)

This is the code for NLPCC 2020 paper Label-Wised Document Pre-Training for Multi-Label Text Classification

Requirements

• Ubuntu 16.04
• Python >= 3.6.0
• PyTorch >= 1.3.0

Reproducibility

• --data and --outputs

We provide the proprecessed RMSC and AAPD datasets and pretrained checkpoints of LW-LSTM+PT+FT model and HLW-LSTM+PT+FT model to make sure reproducibility. Please download from the link and decompress to the root directory of this repository.

--data
    |--aapd
    	|--label_test
    	|--label_train
    	...
    |--rmsc
    	|--rmsc.data.test.json
    	|--rmsc.data.train.json
    	|--rmsc.data.valid.json
    aapd_word2vec.model
    aapd_word2vec.model.wv.vectors.npy
    aapd.meta.json
    aapd.pkl
    rmsc_word2vec.model
    rmsc_word2vec.model.wv.vectors.npy
    rmsc.meta.json
    rmsc.pkl
--outputs
    |--aapd
    |--rmsc

Note that the data/aapdand data/rmsc is the initial dataset. Here we provide a split of RMSC (i.e. RMSC-V2).

• Testing on AAPD

python classification.py -config=aapd.yaml -in=aapd -gpuid [GPU_ID] -test

• Testing on RMSC

python classification.py -config=rmsc.yaml -in=rmsc -gpuid [GPU_ID] -test

Preprocessing

If you want to preprocess the dataset by yourself, just run the following command with name of dataset (e.g. RMSC or AAPD).

PYTHONHASHSEED=1 python preprocess.py -data=[RMSC/AAPD]

Note that PYTHONHASHSEED is used in word2vec.

Pre-Train

Pre-train the LW-PT model.

python pretrain.py -config=[CONFIG_NAME] -out=[OUT_INFIX] -gpuid [GPU_ID] -train -test

• CONFIG_NAME: aapd.yaml or rmsc.yaml
• OUT_INFIX: infix of outputs directory contains logs and checkpoints

MLTC Task

Train the downstream model for MLTC task.

python classification.py -config=[CONFIG_NAME] -in=[IN_INFIX] -out=[OUT_INFIX] -gpuid [GPU_ID] -train -test

• IN_INFIX: infix of inputs directory contains pre-trained checkpoints

Others

• build a static documents representation to facilitate downstream tasks

python build_doc_rep.py -config=[CONFIG_NAME] -in=[IN_INFIX] -gpuid [GPU_ID]

Not used unless necessary.

• make RMSC-V2 dataset: tests/make_rmsc.py
• visual document embeddings: tests/visual_emb.py
• visual labels F1 score: tests/visual_label_f1.py
• case study: tests/case_study.py

Reference

If you consider our work useful, please cite the paper:

@inproceedings{liu2020label,
	title="Label-Wise Document Pre-Training for Multi-Label Text Classification",
	author="Han Liu, Caixia Yuan and Xiaojie Wang",
	booktitle="CCF International Conference on Natural Language Processing and Chinese Computing",
	year="2020"
}

lw-pt's Issues

the download link is expired.

Sorry to interrupt you that the download link is expired.Could you please resent the link？
作者您好！我是一个计算机研究生二年级的学生，非常感谢您把代码开源，但是下载数据的链接失效了，您能不能再发一次呢？