This is the code for our ACL 2021 paper Question Answering over Temporal Knowledge Graphs
Clone and create a conda environment
git clone [email protected]:apoorvumang/CronKGQA.git
cd CronKGQA
conda create --prefix ./cronkgqa_env python=3.8
conda activate ./cronkgqa_env
We use TComplEx KG Embeddings as proposed in Tensor Decompositions for temporal knowledge base completion. We use a slightly modified version of their code from https://github.com/facebookresearch/tkbc
Install tkbc requirements and setup tkbc
conda install --file requirements_tkbc.txt -c pytorch
python setup_tkbc.py install
Install CronKGQA requirements
conda install --file requirements.txt -c conda-forge
Download and unzip data.zip
and models.zip
in the root directory.
Drive: https://drive.google.com/drive/folders/15L4bpGEvCCp7Kuz6xJOFBQFzQGWKJ9rL?usp=sharing, or use gdown
gdown https://drive.google.com/uc\?id\=1wilPf3qohD-6156Daaz5M6GRuJiRP3P4
gdown https://drive.google.com/uc\?id\=18w_aPl-oLfWnTLnoMnTU9Pm4El1T9wkB
unzip -q data.zip && unzip -q models.zip
rm data.zip && rm models.zip
Run a jupyter notebook in the root folder. Make sure to activate the correct environment before running the notebook
The notebook cronkgqa_testing.ipynb
can be used to test a model's responses to any textual question, provided you give the list of entities and times in the question as well - this is needed since perfect entity linking is assumed. You can explore the dataset for questions which have entity annotation and modify those questions. You can also make a reverse dict of data/wikidata_big/kg/wd_id2entity_text.txt
and search for wikidata ids of an entity that you want.
Finally you can run training of QA model using these trained tkbc embeddings. embedkgqa model = cronkgqa (will fix naming etc. soon)
CUDA_VISIBLE_DEVICES=1 python -W ignore ./train_qa_model.py --frozen 1 --eval_k 1 --max_epochs 200 \
--lr 0.00002 --batch_size 250 --mode train --tkbc_model_file tcomplex_17dec.ckpt \
--dataset wikidata_big --valid_freq 3 --model embedkgqa --valid_batch_size 50 \
--save_to temp --lm_frozen 1 --eval_split valid
Evaluating the pretrained model (CronKGQA):
CUDA_VISIBLE_DEVICES=1 python -W ignore ./train_qa_model.py \
--mode eval --tkbc_model_file tcomplex_17dec.ckpt \
--dataset wikidata_big --model embedkgqa --valid_batch_size 50 \
--load_from cronkgqa_29may --eval_split test
Please explore the qa_models.py file for other models, you can change the model by providing the --model parameter.
If you used our work or found it helpful, please use the following citation:
@inproceedings{saxena2021cronkgqa,
title={Question Answering over Temporal Knowledge Graphs},
author={Saxena, Apoorv and Chakrabarti, Soumen and Talukdar, Partha},
booktitle={Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics},
year={2021}
}