Dataset

QA benchmark

1. WebQuestionSP：Same as the original WebQuestionSP QA dataset.
2. CWQ: Same as the original CWQ dataset.

KG

1. Setup Freebase: We use the whole freebase as the knowledge base. Please follow Freebase-Setup to build a Virtuoso for the Freebase dataset.
2. To improve the data accessing efficiency, we extract a 2-hop topic-centric subgraph for each question in WebQSP and a 4-hop topic-centric subgraph for each question in CWQ to create relatively small knowledge graphs. We extract these small knowledge graphs following NSM. You can download the graphs from here.

Running Instructions for WebQSP

Step0: Prepare the weak-supervised dataset for training the retriever：

cd WebQSP Q, run the following scripts.

python run_preprocess.py

Step1: Train the retriever：

python run_train_retriever.py

Step2: Extract a subgraph for each data instance：

python run_retrieve_subgraph.py

Step3: Train the reasoner：

python run_train_nsm.py

Step4: Fine-tune the retriever by the feeback of the reasoner：

python run_retriever_finetune.py

You can also directly run：

./run.sh

Download the data folder tmp from here.

For CWQ, you can run ./cwq/run.sh

If you have any questions about the code, please contact Xiaokang Zhang ([email protected])!