QDTQA

Code for AAAI 2023 research track paper "Question Decomposition Tree for Answering Complex Questions over Knowledge Bases"

Download Paper

Requirements

Knowledge Base

The latest offical data dump of Freebase can be downloaded here.

Python enviroment

• Python >= 3.6
• Pytorch = 1.8.0
• CUDA >= 11.0 (for GPU)
• tensorflow = 2.6.2

To install, run pip install -r requirements.txt in the derectoryClue-Decipher and QDT2Expr. In case of CUDA problems, consult the official PyTorch Get Started guide.

Data and cache

The cache of QDTQA can be found in here

Please put them under the QDT2Expr directory for running QA model.

Run Clue-Decipher

Clue-Decipher is the proposed two-stage decomposition model.

The source code is in QDTQA/Clude-Decipher, hyperparameter is set in config in the run/run_decomp.py

config = {
        'gpu_num': find_free_gpu(0, 20000),
        "max_length": 48,  
        "batch_size": 64,  
        'seed': 42, # random seed
        "epoch": 100,
        'other': '',  
        'action': 'train',  
        'seq_eval': False, # sequence-based or tree-based
    }

Train

The trained model will be saved in model/your_exp_name.pkl

Run run/run_decomp.py with action=train in config

Predict

Predict decomposition file will be saved in data/aaai_out

Run run/run_decomp.py with action=cwq_qa in config

Run QDTQA

QDTQA is the proposed decomposition-based KBQA model, with the decomposition of question obtained by Clue-Decipher as part of the input.

The source code is in QDTQA/QDT2Expr.

Run within QDT2SExpr directoty

Train

Model will be saved in QDT2SExpr/exps/gen_CWQ_nlq_newqdt_candEnt

Please specify the path of your decomposition result obtained from Clue-Decipher in inputDataset/gen_dataset.py,

cwq_load_and_cache_gen_examples()        

if args.new_qdt:
    qdt_file = your_decomposition_result_path        

Then run this scipt

CUDA_VISIBLE_DEVICES=0 sh scripts/run_gen_nlq_newqdt_candEnt.sh train nlq_newqdt_candEnt

Predict

Predict result will be saved in QDT2SExpr/results/gen/CWQ_test_nlq_newqdt_candEnt

CUDA_VISIBLE_DEVICES=0 sh scripts/run_gen_nlq_newqdt_candEnt.sh predict nlq_newqdt_candEnt test

Evaluate

Evaluate result will be saved in QDT2SExpr/results/gen/CWQ_test_nlq_newqdt_candEnt/final_eval_results.txt

python eval_topk_prediction.py --split test --pred_file results/gen/CWQ_test_nlq_newqdt_candEnt/top_k_predictions.json

Run EDGQA with Clue-Decipher

The code for EDGQA can be found in EDGQA.

A modified version of EDGQA that can be run with EDG format decomposition files is available in EDGQA_Exp/EDGQA.

Scripts and resources for converting QDT format to EDG format are available in EDGQA_Exp/QDT2EDG_convert.

Shell scripts for running tests are in EDGQA_Exp/run.

Cite

@article{Huang_Cheng_Shu_Bao_Qu_2023,
  title={Question Decomposition Tree for Answering Complex Questions over Knowledge Bases},
  volume={37},
  url={https://ojs.aaai.org/index.php/AAAI/article/view/26519},
  DOI={10.1609/aaai.v37i11.26519},
  number={11},
  journal={Proceedings of the AAAI Conference on Artificial Intelligence},
  author={Huang, Xiang and Cheng, Sitao and Shu, Yiheng and Bao, Yuheng and Qu, Yuzhong},
  year={2023},
  month={Jun.},
  pages={12924-12932}
}