Inzva AI Projects #4, Text Categorization Using GCN
Our project aims to tackle the text classification problem with novel approaches Graph Convolutional Networks and Graph Attention Networks using Deep Learning algorithms and Natural Language Processing Techniques.
It is applicable for only Linux distros. You can update the commands and use the equivalent ones in other distros (Mac, Windows, etc.) Executing buid.sh will create a new virtual environment in the project folder and install dependencies into that. Run the following command to build:
bash build.sh Be sure that your computer is connected to internet. It can take a while to download and install the dependendencies.
Available Datasets:
- 20ng (Newsgroup Dataset)
- R8 (Reuters News Dataset with 8 labels)
- R52 (Reuters News Dataset with 52 labels)
- ohsumed (Cardiovascular Diseases Abstracts Dataset)
- mr (Movie Reviews Dataset)
- cora (Citation Dataset)
- citeseer (Citation Dataset)
- pubmed (Citation Dataset)
Preprocess:
venv/bin/python3 preprocess.py <DATASET_NAME>Example: venv/bin/python3 preprocess.py R8
Train:
venv/bin/python3 train.py <DATASET_NAME>Example: venv/bin/python3 train.py R8
$ python3 preprocess.py R8
[WARN] directory:'/home/linuxuser/Desktop/GCNN - Repo/gcn_text_categorization/data/corpus.cleaned/' already exists, not overwritten.
[nltk_data] Downloading package stopwords to venv/nltk_data/...
[nltk_data] Unzipping corpora/stopwords.zip.
[INFO] Cleaned-Corpus Dir='/home/linuxuser/Desktop/GCNN - Repo/gcn_text_categorization/data/corpus.cleaned/'
[INFO] Rare-Count=<5>
[INFO] ========= CLEANED DATA: Removed rare & stop-words. =========
[WARN] directory:'/home/linuxuser/Desktop/GCNN - Repo/gcn_text_categorization/data/corpus.shuffled/' already exists, not overwritten.
[WARN] directory:'/home/linuxuser/Desktop/GCNN - Repo/gcn_text_categorization/data/corpus.shuffled/meta/' already exists, not overwritten.
[WARN] directory:'/home/linuxuser/Desktop/GCNN - Repo/gcn_text_categorization/data/corpus.shuffled/split_index/' already exists, not overwritten.
[INFO] Shuffled-Corpus Dir='/home/linuxuser/Desktop/GCNN - Repo/gcn_text_categorization/data/corpus.shuffled/'
[INFO] ========= SHUFFLED DATA: Corpus documents shuffled. =========
[WARN] directory:'/home/linuxuser/Desktop/GCNN - Repo/gcn_text_categorization/data/corpus.shuffled/vocabulary/' already exists, not overwritten.
[WARN] directory:'/home/linuxuser/Desktop/GCNN - Repo/gcn_text_categorization/data/corpus.shuffled/word_vectors/' already exists, not overwritten.
[nltk_data] Downloading package wordnet to venv/nltk_data/...
[nltk_data] Unzipping corpora/wordnet.zip.
[INFO] Vocabulary Dir='/home/linuxuser/Desktop/GCNN - Repo/gcn_text_categorization/data/corpus.shuffled/vocabulary/'
[INFO] Word-Vector Dir='/home/linuxuser/Desktop/GCNN - Repo/gcn_text_categorization/data/corpus.shuffled/word_vectors/'
[INFO] ========= PREPARED WORDS: Vocabulary & word-vectors extracted. =========
[WARN] directory:'/home/linuxuser/Desktop/GCNN - Repo/gcn_text_categorization/data/corpus.shuffled/node_features//R8' already exists, not overwritten.
[INFO] x.shape= (4937, 300), y.shape= (4937, 8)
[INFO] tx.shape= (2189, 300), ty.shape= (2189, 8)
[INFO] allx.shape=(13173, 300), ally.shape=(13173, 8)
[INFO] ========= EXTRACTED NODE FEATURES: x, y, tx, ty, allx, ally. =========
[WARN] directory:'/home/linuxuser/Desktop/GCNN - Repo/gcn_text_categorization/data/corpus.shuffled/adjacency/' already exists, not overwritten.
[INFO] Adjacency Dir='/home/linuxuser/Desktop/GCNN - Repo/gcn_text_categorization/data/corpus.shuffled/adjacency/'
[INFO] ========= EXTRACTED ADJACENCY MATRIX: Heterogenous doc-word adjacency matrix. =========$ python3 train.py R8
[2020/8/16 19:12:00] Epoch:1, train_loss=2.10250, train_acc=0.04436, val_loss=2.06262, val_acc=0.62044, time=4.47397
[2020/8/16 19:12:02] Epoch:2, train_loss=1.91917, train_acc=0.66235, val_loss=2.05057, val_acc=0.68978, time=2.37341
[2020/8/16 19:12:05] Epoch:3, train_loss=1.80157, train_acc=0.72514, val_loss=2.04366, val_acc=0.72263, time=2.30123
[2020/8/16 19:12:07] Epoch:4, train_loss=1.73192, train_acc=0.76727, val_loss=2.03955, val_acc=0.75365, time=2.29406
...
[2020/8/16 19:14:40] Epoch:61, train_loss=1.42440, train_acc=0.99149, val_loss=2.01033, val_acc=0.96350, time=2.95969
[2020/8/16 19:14:43] Epoch:62, train_loss=1.42407, train_acc=0.99190, val_loss=2.01033, val_acc=0.96168, time=2.79453
[2020/8/16 19:14:45] Epoch:63, train_loss=1.42377, train_acc=0.99271, val_loss=2.01034, val_acc=0.96168, time=2.56621
[2020/8/16 19:14:45] Early stopping...
[2020/8/16 19:14:45] Optimization Finished!
[2020/8/16 19:14:46] Test set results:
[2020/8/16 19:14:46] loss= 1.79863, accuracy= 0.97076, time= 0.85016
[2020/8/16 19:14:46] Test Precision, Recall and F1-Score...
[2020/8/16 19:14:47] precision recall f1-score support
[2020/8/16 19:14:47]
[2020/8/16 19:14:47] 0 0.9826 0.9917 0.9871 1083
[2020/8/16 19:14:47] 1 0.9706 0.8148 0.8859 81
[2020/8/16 19:14:47] 2 0.9826 0.9741 0.9784 696
[2020/8/16 19:14:47] 3 0.9435 0.9669 0.9551 121
[2020/8/16 19:14:47] 4 0.8454 0.9425 0.8913 87
[2020/8/16 19:14:47] 5 0.9012 0.9733 0.9359 75
[2020/8/16 19:14:47] 6 0.9615 0.6944 0.8065 36
[2020/8/16 19:14:47] 7 1.0000 1.0000 1.0000 10
[2020/8/16 19:14:47]
[2020/8/16 19:14:47] accuracy 0.9708 2189
[2020/8/16 19:14:47] macro avg 0.9484 0.9197 0.9300 2189
[2020/8/16 19:14:47] weighted avg 0.9715 0.9708 0.9703 2189
[2020/8/16 19:14:47]
[2020/8/16 19:14:47] Macro average Test Precision, Recall and F1-Score...
[2020/8/16 19:14:47] (0.9484369779553932, 0.9197363948390138, 0.9300186011259608, None)
[2020/8/16 19:14:47] Micro average Test Precision, Recall and F1-Score...
[2020/8/16 19:14:48] (0.9707629054362723, 0.9707629054362723, 0.9707629054362723, None)
[2020/8/16 19:14:48] Embeddings:
Word_embeddings:7688
Train_doc_embeddings:5485
Test_doc_embeddings:2189
Word_embeddings::48]
[[0. 0.16006473 0. ... 0.23560655 0.03900092 0.03204489]
[0.77342826 0. 0. ... 0. 0.5302985 0.6264921 ]
[0.7280124 0.2820953 0. ... 0. 0.11174901 0.5485109 ]
...
[0.3487121 0.00971795 0.16308378 ... 0.23769674 0.19000074 0.06477314]
[0.04132688 0.10753749 0.06603149 ... 0.05606151 0.13652363 0.0443929 ]
[0.06743763 0.23347118 0.36717737 ... 0.15584956 0. 0. ]]- [Liang Yao, Chengsheng Mao, Yuan Luo, 2018] Graph Convolutional Networks for Text Classification
- [Kipf and Welling, 2017] Semi-supervised Classification with Graph Convolutional Networks
- [Ankit Pal, 2020] Multi-label Text Classification using Attention based Graph Neural Network
- [Petar Velickovic, 2017] Graph Attention Networks
- PyTorch Implementation of Graph Attention Networks: https://github.com/Diego999/pyGAT
- PyTorch implementation of Graph Convolutional Networks: https://github.com/tkipf/pygcn