The repository which contains the code and pre-trained models for our paper: Explanation Graph Generation via Generative Pre-training over Synthetic Graphs.(ACL2023-Findings)

In this paper, we present $EG^3P$ (for Explanation Graph Generation via Generative Pre-training over Synthetic Graphs), a pre-training structure for explanation graph generation. As shown in the Figure 1, $EG^3P$ contains a pre-training method in the form of "text-to-graph", and an automated process for automatically synthesizing aligning corpus.

In the process of pre-training, we input the synthesized query and simulated knowledge base, and the model will output the reasoning graph. In addition, we have constructed a large number of composite graphs and natural language queries based on external structured knowledge bases, and the specific synthesis process is shown in the following Figure 2.

conda creative -n EG3P python=3.8 
pip install -r requirement.txt

You could check the pre-training data and pre-trained model in the following link: EG3P_Data_Model

For detailed info, you could check the README.md in the dir of data.

• Process the data

bash process_dataset.sh DATASET_PATH MODEL_DIR_PATH

where DATASET_PATH is the path of the dataset path(.src/.tgt), and MODEL_DIR_PATH is the path of the model dir.

• Pretraining

bash pretrain.sh DATASET_PATH MODEL_PATH

where DATASET_PATH is the path of the dataset path(processed), and MODEL_PATH is the path of the model file(xx.pt)

• Fine-tuning

bash finetune.sh TASK_NAME DATASET_PATH MODEL_NAME MODEL_PATH 

In the script:

• TASK_NAME: The name of downstream task: ExplaGraphs / CSQA / OBQA
• DATASET_PATH: The path of the dataset path(processed)
• MODEL_NAME: The name of the pre-trained model. (only used for naming the checkpoint_dir)
• MODEL_PATH: The path of the model file(xx.pt)

bash eval.sh DATASET_DIR MODEL_PATH PREDICT_DIR

In this script:

• DATASET_DIR: The path of the dataset path ( processed, and containing the test.src )
• MODEL_PATH: The path of the model file(xx.pt)
• PREDICT_DIR: The path of the result of output