Semantic Structure based Query Graph Prediction for Question Answering over Knowledge Graph [COLING 2022]

This is the source code of SSKGQA. For the full project, including MetaQA and WSP datasets, please refer to Google Drive.

1) Overview

The architecture of our proposed SSKGQA is depicted in the diagram below. Given a question q , we assume the topic entity of q has been obtained by preprocessing. Then the answer to q is generated by the following steps.

Step 1: The semantic structure of q is predicted by a novel Structure-BERT classifier. For the example above, q is a 2-hop question and the classifier predicts its semantic structure as SS2.

Step 2: We retrieve all the candidate query graphs (CQGs) of q by enumeration, and use the predicted semantic structure SS2 as the constraint to filter out noisy candidate query graphs and keep the candidates with correct structure (CQG-CS). Afterwards, a BERT-based ranking model is used to score each candidate query graph in CQG-CS, and the top-1 highest scored candidate is selected as the query graph g for question q. Finally, the selected query graph is issued to KG to retrieve the answer Sergei Kozlov.

2) Configuration

Step 1:

1. Python>=3.6.5

2. Transformer: pip install transformers==3.0.0

3. Torch with 1.5.0: pip install torch==1.5.0+cu101 torchvision==0.6.0+cu101 -f https://download.pytorch.org/whl/torch_stable.html

4. Latest Tensorflow: pip install Tensorflow

Step 2:

1. Python>=3.6.5

2. Transformer: pip install transformers==4.3.3

3. torch==1.7.1: pip install torch==1.7.1+cu101 torchvision==0.8.2+cu101 torchaudio==0.7.2 -f https://download.pytorch.org/whl/torch_stable.html:

3) How to run

For excample:

the training file in step 1 and step2 all are train.py, just run it by
CUDA_VISIBLE_DEVICES=1 python train.py,  please use the default parameters.

4) Demo display

Question: what is my timezone in louisiana ?

ground truth query graph is: louisiana (topic entity)#location.location.time_zones (relation)#?x (the answer which to be got)#1 (SS1 for our proposed Semantic structure)

candidate query graph:

louisiana#base.biblioness.bibs_location.country#?x#1
louisiana#location.location.events#?x#1
louisiana#location.location.partially_contains#?x#1

-----louisiana#location.location.time_zones#?x#1-------

louisiana#government.political_district.representatives#?x#1
louisiana#location.administrative_division.country#?x#1
louisiana#base.aareas.schema.administrative_area.administrative_parent#?x#1
louisiana#location.administrative_division.first_level_division_of#?x#1
louisiana#location.location.partiallycontains#?x#1
louisiana#book.book_subject.works#?x#1

Prediction results:

5) Citation

Please cite this paper if you find the paper or code is useful.

@article{li2022sskgqa,
  title={Semantic Structure based Query Graph Prediction for Question Answering over Knowledge Graph},
  author={Li, Mingchen and Ji, Shihao},
  journal={International Conference on Computational Linguistics (COLING)},
  year={2022}
}