PaSIG

The implementation of PaSIG in the paper:

Shiyu Wang, Gang Zhou, Jicang Lu, Jing Chen, Ningbo Huang. "Pre-trained Semantic Interaction based Inductive Graph Neural Networks for Text Classification". In 31st International Conference on Computational Linguistics (COLING 2025).

File Trees

.
├── data
│     ├── embed_bert
│     ├── graph
│     ├── raw
│     │     ├── mr.labels.txt
│     │     ├── mr.texts.txt
│     │     ├── ...
│     │     ├── R52.labels.txt
│     │     └── R52.texts.txt
│     └── temp
├── log
├── out
├── proc
│     ├── dataset_config.py
│     ├── encoder_bert.py
│     ├── graph_builder.py
│     ├── my_tfidf.py
│     └── preprocess_data.py
├── ptm
│     └── bert-base-uncased
│         ├── config.json
│         ├── pytorch_model.bin
│         ├── tokenizer_config.json
│         ├── tokenizer.json
│         └── vocab.txt
└── src
    ├── bert_model.py
    ├── dataset_graph_batch.py
    ├── dataset_graph.py
    ├── dataset_text.py
    ├── gnn_layer.py
    ├── gnn_model_batch.py
    ├── gnn_model.py
    ├── train_bert.py
    ├── train_gnn_batch.py
    ├── train_gnn.py
    └── utils.py

Usage

1. Prepare the raw data

Put the raw files in path data/raw/, including raw text file (*.texts.txt) and label file (*.labels.txt).

2. Preprocess the data

Set dataset parameters in the file proc/dataset_config.py and change the selected dataset name.

Run the preprocessing code proc/preprocess_data.py and proc/graph_builder.py.

  python proc/preprocess_data.py

  python proc/graph_builder.py

The intermediate data file will be saved in the path data/temp/ and data/graph/.

3. Download BERT model

Download Bert-Base-Uncased model and unzip it in the path ptm/.

4. Train BERT model to obtain text embedding

• Train BERT model, run the code src/train_bert.py. For example:

  python src/train_bert.py --dataset mr --epochs 20

• Get texts and words encoding by BERT, run the code proc/encoder_bert.py. (You also need to change the Dataset name)

  python proc/encoder_bert.py

5. Train GNN model and observe node embedding

• Train PaSIG, run the code src/train_gnn.py. For example:

  python src/train_gnn.py --dataset mr --gnn_model gfus

• Train PaSIG-S, run the code src/train_gnn_batch.py. For example:

  python src/train_gnn_batch.py --dataset mr --gnn_model gfus

  You can choose 4 gnn components to run: gcn, gin, sage, gfus.