Text GCN Tutorial

This tutorial (currently under improvement) is based on the implementation of Text GCN in our paper:

Liang Yao, Chengsheng Mao, Yuan Luo. "Graph Convolutional Networks for Text Classification." In 33rd AAAI Conference on Artificial Intelligence (AAAI-19)

Require

Python 2.7 or 3.6

Tensorflow >= 1.4.0

Example input data

The Ohsumed corpus is from the MEDLINE database, which is a bibliographic database of important medical literature maintained by the National Library of Medicine

In this tutorial, we created a subsample of the 2,762 unique diseases abstracts from 3 categories

• C04: Neoplasms
• C10: Nervous System Diseases
• C14: Cardiovascular Diseases

As we focus on single-label text classification, the documents belonging to multiple categories are excluded

1230 train (use 10% as validation), 1532 test

1. /data/ohsumed_3.txt indicates document names, training/test split, document labels. Each line is for a document.

2. /data/corpus/ohsumed_3.txt contains raw text of each document, each line is for the corresponding line in /data/ohsumed_3.txt

Reproduing Results

1. Run python remove_words.py ohsumed_3

2. Run python build_graph.py ohsumed_3

3. Run python train.py ohsumed_3

Example output

2019-04-04 22:58:26.244395: I tensorflow/core/platform/cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: SSE4.1 SSE4.2 AVX AVX2 FMA
Epoch: 0001 train_loss= 1.09856 train_acc= 0.41463 val_loss= 1.08209 val_acc= 0.48780 time= 29.13731
Epoch: 0002 train_loss= 1.08044 train_acc= 0.49865 val_loss= 1.05469 val_acc= 0.47967 time= 23.00088
Epoch: 0003 train_loss= 1.05075 train_acc= 0.49865 val_loss= 1.02113 val_acc= 0.47967 time= 21.82401
Epoch: 0004 train_loss= 1.01430 train_acc= 0.49955 val_loss= 0.98582 val_acc= 0.48780 time= 21.42816
Epoch: 0005 train_loss= 0.97174 train_acc= 0.50678 val_loss= 0.95375 val_acc= 0.51220 time= 21.44958
Epoch: 0006 train_loss= 0.93406 train_acc= 0.51220 val_loss= 0.92789 val_acc= 0.55285 time= 24.01502
......
Epoch: 0074 train_loss= 0.01921 train_acc= 0.99819 val_loss= 0.09674 val_acc= 0.96748 time= 24.01229
Epoch: 0075 train_loss= 0.02093 train_acc= 0.99909 val_loss= 0.09715 val_acc= 0.96748 time= 24.08436
Early stopping...
Optimization Finished!
Test set results: cost= 0.24295 accuracy= 0.92167 time= 7.60145
10456
Test Precision, Recall and F1-Score...
             precision    recall  f1-score   support

          0     0.8882    0.8363    0.8614       342
          1     0.9438    0.9517    0.9477       600
          2     0.9174    0.9407    0.9289       590

avg / total     0.9212    0.9217    0.9212      1532

Visualizing Documents

Run python tsne.py

Example Visualization