This repository is for implementation for the paper "EFOk-CQA: Towards Knowledge Graph Complex Query Answering beyond Set Operation".
We have utilized a CSP solver provided in the python-constraint package, please install it by:
pip install python-constraint
We have also utilized the pytorch-geometric and networkx package, please install it by:
conda install pyg -c pyg
conda install networkx
Please download the EFOk-CQA dataset from here,
the data of three knowledge graphs can be downloaded separately and put it in the data folder, as well as a file named
DNF_EFO2_23_4123166.csv which is used to store the abstract query graph(query type) for the EFOX experiment.
The DNF_EFO2_23_4123166.csv should be put into the data folder.
Then, after unzipping the query data. an example data folder should look like this:
data/FB15k-237-EFOX-final/
- kgindex.json
- train_kg.tsv
- valid_kg.tsv
- test_kg.tsv
- test-type0000-EFOX-qaa.json
- test-type0001-EFOX-qaa.json
- ......
where the test-type0000-EFOX-qaa.json is used for the EFOX experiment, containing the data for query type0000.
The kgindex.json and train_kg.tsv are the index file and the training graph for the knowledge graph respectively,
the valid_kg.tsv and test_kg.tsv are the validation graph and the test graph respectively. They are used for data generation.
In each test-type0000-EFOX-qaa.json, it contains a dict with the key be the formula and the value is a list of query:
{formula: [query1, query2, ...]}.
To reproduce the experiment in the paper, we have provided the checkpoint for each model foreach knowledge graph, we offer the checkpoint for six representative model (BetaE, LogicE, ConE, CQD, LMPNN, FIT), which can be downloaded from here,
It should be unzipped and put in the ckpt folder.
An example of the ckpt sub folder, which includes the model trained on the knowledge graph ``FB15k-237'' should look like this:
ckpt/FB15k-237
- BetaE_full/checkpoint
- LogicE_full/450000.ckpt
- ConE_full/300000.ckpt
- CQD/FB15k-237-model-rank-1000-epoch-100-1602508358.pt
- LMPNN/lmpnn-FB15k-237.ckpt
- FIT/torch_0.005_0.001.ckpt
where each sub folder is the checkpoint for each model, and the name of the sub folder is the name of the model.
We have the powerful frame that supports several key functionalities for the task of complex query answering, you can also sample the query by yourself following the instruction.
If you have downloaded the EFOk-CQA dataset, you can also skip this section.
To try to enumerate the abstract query graph, please run the following command:
python data_preparation/create_qg.py
it should output the abstract query graph in the data folder, with the name of DNF_EFO2_23_4123166.csv_filtered.csv,
which is used to store the abstract query graph(query type) for the EFOX experiment. You can also change hyperparameters used
in this code to explore other possible combinatorial space different from the one in the paper.
Ground the abstract query graph to become a query graph requires two functionalities:
- Ground entities and relations,
- Compute the answer for the grounded query graph.
We give a example of sampling the query graph for the knowledge graph FB15k-237, where each type of query has 1000 samples if it does not negation and 500 otherwise, please run the following command:
python data_preparation/sample_query.py --output_folder data/FB15k-237-EFOX-final --data_folder data/FB15k-237-EFOX-final --num_positive 1000 --num_negative 500
For query embedding method, including BetaE, LogicE, ConE, please run the following command:
python QG_EFOX.py --config config/LogicE_FB15k-237_EFOX.yaml
which is an example for LogicE method on FB15k-237 dataset. The config file in the config folder is used to specify
the model and knowledge graph used in the experiment.
For LMPNN, note that you need to download the CQD checkpoint as well checkpoint of LMPNN since LMPNN is built upon CQD, please run the following command for LMPNN on FB15k-237 dataset:
python evaluate_lmpnn.py \
--task_folder data/FB15k-237-EFOX-final \
--checkpoint_path ckpt/FB15k-237/CQD/FB15k-237-model-rank-1000-epoch-100-1602508358.pt \
--checkpoint_path_lmpnn ckpt/FB15k-237/LMPNN/lmpnn-FB15K-237.ckpt \
--embedding_dim 1000
For LMPNN on FB15k:
python evaluate_lmpnn.py \
--task_folder data/FB15k-EFOX-final \
--checkpoint_path ckpt/FB15k/CQD/FB15k-model-rank-1000-epoch-100-1602520745.pt \
--checkpoint_path_lmpnn ckpt/FB15k/LMPNN/lmpnn-FB15K.ckpt \
--hidden_dim 8192
For LMPNN on NELL:
python3 evaluate_lmpnn.py \
--task_folder data/NELL-EFOX-final \
--checkpoint_path ckpt/NELL/CQD/NELL-model-rank-1000-epoch-100-1602499096.pt \
--checkpoint_path_lmpnn ckpt/NELL/LMPNN/lmpnn-NELL.ckpt \
--hidden_dim 8192 \
--temp 0.1
For CQD on FB15k-237:
python evaluate_lmpnn.py \
--task_folder data/FB15k-237-EFOX-final \
--reasoner gradient \
--checkpoint_path ckpt/FB15k-237/CQD/FB15k-237-model-rank-1000-epoch-100-1602508358.pt
For CQD on FB15k:
python evaluate_lmpnn.py \
--task_folder data/FB15k-EFOX-final \
--reasoner gradient \
--checkpoint_path ckpt/FB15k/CQD/FB15k-model-rank-1000-epoch-100-1602520745.pt \
--hidden_dim 8192
For CQD on NELL:
python3 evaluate_lmpnn.py \
--task_folder data/NELL-EFOX-final \
--reasoner gradient \
--checkpoint_path ckpt/NELL/CQD/NELL-model-rank-1000-epoch-100-1602499096.pt \
--hidden_dim 8192
For FIT, please run the following command to run the expriment on FB15k-237:
python solve_EFOX.py
If you want to try to use FIT on KB15k or NELL, please run the following command:
python solve_EFOX.py --ckpt ckpt/FB15k/FIT/torch_0.005_0.001.ckpt --data_folder data/FB15k-EFOX-final
python solve_EFOX.py --ckpt ckpt/NELL/FIT/torch_0.0002_0.001.ckpt --data_folder data/NELL-EFOX-final
We note it may encounter out-of-memory error when running the FIT model on KB15k and NELL as mentioned in the paper, which indicates that FIT face the challenge of scalability.
As there are numerous abstract query graphs (query types), we can aggregate the result of each query type to get the final result.
Please create a folder to record the benchmark result, just like the following:
result/FB15k-237_result
-BetaE_test
-LogicE_test
-ConE_test
-CQD_test
-LMPNN_test
-FIT_test
Then run the following code and get the presentation of the table, for knowledge graph FB15k-237, queries with one free variable:
python construct_and_analyze.py --out_folder result --dataset FB15k-237 --model LogicE --variable 1 --construct 1
The --variable is used to specify the number of variable in the query graph, it can be set to 1 or 2.