Source code for "Question-directed Reasoning with Relation-aware Graph Attention Network for Complex Question Answering over Knowledge Graph"
There are two datasets used in this work, WebQuestionsSP and ComplexWebQuestions. Preprocessed data can be directly accessed in this link (~1.05GB).
All the training logs for the experiments in the paper and the corresponding model checkpoints can be accesses in this link (~3.14GB).
All the commands are listed in the scripts directory. Before rerunning the training process, please make sure the pretrained data are downloaded and untared into datasets directory; before doing the inference, please make sure the model checkpoints are downloaded and untared into cache directory.
Followers can rerun the training process by the shell scripts with train in their name, or reproduce the experiment results by scripts with inference in their name.
git clone https://github.com/zxgx/QRGAT.git
cd QRGAT
# download the preprocessed data
tar -zxf preprocessed_data.tgz -C <data_dir>
cd datasets
ln -s <data_dir>/CWQ CWQ
ln -s <data_dir>/webqsp webqsp
cd ..
# download the model checkpoints
tar -zxf qrgat.tgz -C cache
bash scripts/<any shell script>
To get the results with entity linking (EL), we use the EL results from RnG-KBQA for WebQSP.
The scripts to annotate our datasets with entity linking are in datasets dir.
To run this setup, please (1) replace the dataset softlink (explained below) with the downloaded dataset and (2) use --enable_entity_linking
If you find this repository helpful, kindly cite:
@article{zhang2024question,
title={Question-Directed Reasoning With Relation-Aware Graph Attention Network for Complex Question Answering Over Knowledge Graph},
author={Zhang, Geng and Liu, Jin and Zhou, Guangyou and Zhao, Kunsong and Xie, Zhiwen and Huang, Bo},
journal={IEEE/ACM Transactions on Audio, Speech, and Language Processing},
year={2024},
publisher={IEEE}
}