KBQA-GST

This is a pytorch implementation of the KBQA-GST model in the paper "Knowledge Base Question Answering with Topic Units"

Dependency

The model is tested in python 3.7 and pytorch 1.3.0

conda create -n kbqagst python=3.7
source activate kbqagst
pip install -r requirements.txt

Prepare Data

First, create a folder data/ under the current directory. Download the GloVe from here as

data/glove.840B.300d.zip

Then, download and merge the pre-processed datasets (ComplexWebQuestion and WebQuestionsSP) from here under the data/ directory. The data contains questions, answers, results of topic entity linking, unit linking and the candidate paths based on the entity linking, unit linking respectively.

Download Pre-trained models

Download the pre-trained models from here. Put the saved-model/ under the current directory.

Run the Pre-trained model

To run the pre-trained models, run python code/RL_Runner.py --task 1 for WebQuestionsSP and python code/RL_Runner.py --task 3 for ComplexWebQuestion, respectively. You can change the argument (deactivate training flag or change the model's name) of the loaded argument in the code/RL_Runner.py file.
To generate the final prediction, change the argument in the file code/GenerateFinalPredictions.py and run python code/GenerateFinalPredictions.py.
To evaluate the WebQuestionsSP via the official evaluation file, run python code/eval.py.

Run a New Model

To run a new model, set the argument of the loaded argument in the code/RL_Runner.py file run python code/RL_Runner.py --task 0.

Please cite the papers if you use our data and code.

@inproceedings{ijcai2019-701,
  title     = {Knowledge Base Question Answering with Topic Units},
  author    = {Lan, Yunshi and Wang, Shuohang and Jiang, Jing},
  booktitle = {Proceedings of the Twenty-Eighth International Joint Conference on
               Artificial Intelligence, {IJCAI-19}},
  publisher = {International Joint Conferences on Artificial Intelligence Organization},             
  pages     = {5046--5052},
  year      = {2019},
  month     = {7},
  doi       = {10.24963/ijcai.2019/701},
  url       = {https://doi.org/10.24963/ijcai.2019/701},
}

Contact Yunshi Lan for any question.