bahuia/hgnet

Data and Code for paper Outlining and Filling: Hierarchical Query Graph Generation for Answering Complex Questions over Knowledge Graphs accepted by IEEE Transactions on Knowledge and Data Engineering, is available for research purposes.

Results

We apply three KGQA benchmarks to evaluate our approach, ComplexWebQuestions (Talmor and Berant, 2018), LC-QuAD (Trivedi et al., 2017), and WebQSP (Yih et al., 2016).

Dataset	Structure Acc.	Query Graph Acc.	Precision	Recall	F1-score	Hit@1
ComplexWebQuestions	66.96	51.68	65.27	68.44	64.95	65.25
ComplexWebQuestions (Bert-base)	72.88	57.80	68.89	73.30	68.88	68.80
LC-QuAD	78.00	60.90	75.82	75.22	75.10	76.00
LC-QuAD (Bert-base)	80.70	63.50	78.92	78.14	78.13	78.70
WebQSP	79.91	62.63	70.22	74.38	70.61	70.37
WebQSP (Bert-base)	85.03	70.74	76.66	79.28	76.62	76.92

Data

• Download and unzip our preprocessed data to ./, you can also running our scripts under ./preprocess to obtain them again.

• Download GloVe Embedding glove.42B.300d.txt and put it to your_glove_path.

• Download our vocabulary from here. Unzip and put it under ./. It contains our used SPARQL cache for Execution-Guided strategy.

Virtuoso SPARQL Service

The two KGs we use contain only triples in English by removing other languages. Please download the KGs virtuoso-opensource we use for the SPARQL query service for the downloaded Freebase and DBpedia. Here is a tutorial on how to install Virtuoso and import the knowledge graph into it.

1. Get root access
2. Edit virtuoso-opensource/database/virtuoso.ini and set the property "DirsAllowed" to your path.
3. Execute the following commands to start the service.

cd virtuoso-opensource/database/
../bin/virtuoso-t -fd

Running Code

1. Training for HGNet

Before training, first set the following hyperparameter in main_train_cwq.sh, main_train_lcq.sh, and main_train_wsp.sh.

--glove_path your_glove_path

Execute the following command for training model on ComplexWebQuestions.

sh train_cwq.sh

Execute the following command for training model on LC-QuAD.

sh train_lcq.sh

Execute the following command for training model on WebQSP.

sh train_wsp.sh

The trained model file is saved under ./runs directory.
The path format of the trained model is ./runs/RUN_ID/checkpoints/best_snapshot_epoch_xx_best_val_acc_xx_model.pt.

2. Training for HGNet with Bert-base

Execute the following command for training model on ComplexWebQuestions.

sh train_plm_cwq.sh

Execute the following command for training model on LC-QuAD.

sh train_plm_lcq.sh

Execute the following command for training model on WebQSP.

sh train_plm_wsp.sh

The trained model file is saved under ./runs directory.
The path format of the trained model is ./runs/RUN_ID/checkpoints/best_snapshot_epoch_xx_best_val_acc_xx_model.pt.

3. Testing for HGNet

Before testing, need to train a model first and set the following hyperparameters in main_eval_cwq.sh, main_eval_lcq.sh, and main_eval_wsp.sh.

--cpt your_trained_model_path
--kb_endpoint your_sparql_service_ip

You can also directly download our trained models from here. Unzip and put it under ./.

Execute the following command for testing the model on ComplexWebQuestions.

sh eval_cwq.sh

Execute the following command for testing the model on LC-QuAD.

sh eval_lcq.sh

Execute the following command for testing the model on WebQSP.

sh eval_wsp.sh

4. Testing for HGNet with Bert-base

Before testing, need to train a model first and set the following hyperparameters in main_eval_plm_cwq.sh, main_eval_plm_lcq.sh, and main_eval_plm_wsp.sh.

--cpt your_trained_model_path
--kb_endpoint your_sparql_service_ip

You can also directly download our trained models from here. Unzip and put it under ./.

Execute the following command for testing the model on ComplexWebQuestions.

sh eval_plm_cwq.sh

Execute the following command for testing the model on LC-QuAD.

sh eval_plm_lcq.sh

Execute the following command for testing the model on WebQSP.

sh eval_plm_wsp.sh

hgnet's Issues

Data link error

Dear Authors,
When I tried to download the data, such as the SPARQL cache, each OneDrive link shows error as above. Will you encounter such a situation? Or is it because the links are expired? Thank you very much.

cannot get your data

Sorry, the link you shared seems to be broken and I can't access your data.

When I execute "eval_plm_wsp. sh" and "eval_wsp. sh", no results returned.

I use the latest version and modify the kb_endpoint in the eval_plm_cwq.sh file and eval_cwq.sh file, but the results are still empty. If i add "--not_use_eg", they work fine

When I execute "main_eval_plm_wsp. sh", the results are all 0.

When I execute "main_eval_wsp. sh", it works fine.

The same problem occurs when I execute "main_eval_plm_cwq. sh".