/TextLevelGNN-DimarzioBian

An unofficial PyTorch implementation for the TextLevelGNN [EMNLP'19] "Text Level Graph Neural Network for Text Classification"

Primary LanguagePython

TextLevelGNN

This is an unofficial PyTorch implementation for the EMNLP'19 paper Text Level Graph Neural Network for Text Classification

Training

Currently support DATASET_NAME: R8, R52, Ohsumed, MR. Preprocessed corpus and GloVe are saved in .pkl files under ./data/.

To be noticed, set args.max_len_text to 150 for Ohsumed, and 100 for all others.

Dataset

If you wanna re-preprocess the data, please read below.

Please refer to textGCN and copy the R8, R52, mr, ohsumed_single_23 folder into data folder.

For word embeddings, please refer to GloVe, download glove.6B.zip, unzip and move the file glove.6B.300d.txt under ./data/ folder.

Results

We are only able to implement similar result on R8 and R52 dataset, while Ohsumed perform quite difference as compared to paper's one.

Accuracy R8 R52 Ohsumed
Test 98.31% 94.55% 67.15%

With the hyperparameter settings that achieves this result

Hyperparameter R8 R52 Ohsumed
batch_size 100 100 100
max_len_text 100 100 100
dropout 0 0 0
Use layer_norm False True True
Use ReLU False False True
Use mean_reduction False False False
Use pretrained False False False
n_degree 2 11 11