This is the code for the ACL 2020 Paper: Every Document Owns Its Structure: Inductive Text Classification via Graph Neural Networks.
Download pre-trained word embeddings glove.6B.300d.txt from here and unzip to the repository.
Build graphs from the datasets in data/corpus/ as:
python build_graph.py [DATASET] [WINSIZE]
Provided datasets include mr,ohsumed,R8andR52. The default sliding window size is 3.
To use your own dataset, put the text file under data/corpus/ and the label file under data/ as other datasets do. Preprocess the text by running remove_words.py before building the graphs.
Start training and inference as:
python train.py [--dataset DATASET] [--learning_rate LR]
[--epochs EPOCHS] [--batch_size BATCHSIZE]
[--hidden HIDDEN] [--steps STEPS]
[--dropout DROPOUT] [--weight_decay WD]
To reproduce the result, large hidden size and batch size are suggested as long as your memory allows. We report our result based on 96 hidden size with 1 batch. For the sake of memory efficiency, you may change according to your hardware.
- Python 3.6+
- Tensorflow/Tensorflow-gpu 1.12.0
- Scipy 1.5.1
Please cite our paper if you use the code:
@inproceedings{zhang-etal-2020-every,
title = "Every Document Owns Its Structure: Inductive Text Classification via Graph Neural Networks",
author = "Zhang, Yufeng and
Yu, Xueli and
Cui, Zeyu and
Wu, Shu and
Wen, Zhongzhen and
Wang, Liang",
editor = "Jurafsky, Dan and
Chai, Joyce and
Schluter, Natalie and
Tetreault, Joel",
booktitle = "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics",
month = jul,
year = "2020",
address = "Online",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2020.acl-main.31",
doi = "10.18653/v1/2020.acl-main.31",
pages = "334--339",
abstract = "Text classification is fundamental in natural language processing (NLP) and Graph Neural Networks (GNN) are recently applied in this task. However, the existing graph-based works can neither capture the contextual word relationships within each document nor fulfil the inductive learning of new words. Therefore in this work, to overcome such problems, we propose TextING for inductive text classification via GNN. We first build individual graphs for each document and then use GNN to learn the fine-grained word representations based on their local structure, which can also effectively produce embeddings for unseen words in the new document. Finally, the word nodes are aggregated as the document embedding. Extensive experiments on four benchmark datasets show that our method outperforms state-of-the-art text classification methods.",
}
Some functions are based on Text GCN. Thank for their work.