日本語

XML-CNN

Implementation of Deep Learning for Extreme Multi-label Text Classification using PyTorch.

Liu, J., Chang, W.-C., Wu, Y. and Yang, Y.: Deep learning fo extreme multi-label text classification, in Proc. of the 40th International ACM SIGIR Conference on Research and Development in Information Retrieval, pp. 115-124 (2017).

Requirements

• Python: 3.6.10 or higher
• PyTorch: 1.6.0 or higher
• torchtext: 0.6.0 or higher
• Optuna: 2.0.0 or higher

You can create a virtual environment of Anaconda from requirements.yml.
We can't guarantee the operation of Anaconda environment other than the one created with requirements.yml.

$ conda env create -f requirements.yml

Datasets

The dataset must be in the same format as it attached to this program.

It contains one document per line.
It's stored in the order of ID, label, and text, separated by TAB from the left side.

{id}<TAB>{labels}<TAB>{texts}

You can get the tokenized RCV1 dataset from Lewis et al. by using this program's bundled data/get_rcv1.py.
Caution: This dataset is tokenized differently than the one used by Liu et al.
Caution: If you want to use this dataset, please read the terms of use (Legal Issues) of the distribution destination.

Lewis, D. D.; Yang, Y.; Rose, T.; and Li, F. RCV1: A New Benchmark Collection for Text Categorization Research. Journal of Machine Learning Research, 5:361-397, 2004. http://www.jmlr.org/papers/volume5/lewis04a/lewis04a.pdf.

Lewis, D. D. RCV1-v2/LYRL2004: The LYRL2004 Distribution of the RCV1-v2 Text Categorization Test Collection (12-Apr-2004 Version). http://www.jmlr.org/papers/volume5/lewis04a/lyrl2004_rcv1v2_README.htm.

Dynamic Max Pooling

This program implements Dynamic Max Pooling based on the method by Liu et al.

The p shown in that paper becomes the d_max_pool_p in ./params.yml.
As in the paper, d_max_pool_p must be a divisible number for the output vector after convolution.

Evaluation Metrics

Precision@K and F1-Score are available for this program.
You can change it from ./params.yml.

How to run

When running at first

Donwload RCV1

Donwload datasets from http://www.ai.mit.edu/projects/jmlr/papers/volume5/lewis04a/lyrl2004_rcv1v2_README.htm.
Caution: This dataset is tokenized differently than the one used by Liu et al.
Caution: If you want to use this dataset, please read the terms of use (Legal Issues) of the distribution destination.

$ cd data
$ python get_rcv1.py

Make valid dataset

$ python make_valid.py train_org.txt

Make dataset for Params Search

$ python make4search.py train_org.txt

Run

$ python train.py

Normal Training

$ python train.py

Params Search

$ python train.py --params_search

or

$ python train.py -s

Force to use cpu

$ python train.py --use_cpu

Acknowledgment

This program is based on the following repositories. Thank you very much for their accomplishments.

• siddsax/XML-CNN (MIT License)
• PaulAlbert31/LabelNoiseCorrection (MIT License)