Implementations for paper "Enhancing SPARQL Query Generation for KBQA Systems by Learning to Correct Triplets".
Create an environment using Python 3.7, and install the dependencies with
pip install -r requirements.txt
The pre-processed datasets that can be used directly for training has been placed under the folder transform/transformers_cache/downloads. And we recommend you to use them.
- LC-QuAD2.0-master: Processed LC-QuAD 2.0 dataset for fine-tuning;
- LC-QuAD2.0-pre: Processed LC-QuAD 2.0 dataset for pre-training;
- QALD_9_PULS: Processed Qald-9-plus dataset for fine-tuning;
- QALD_10: Processed Qald-10 dataset for fine-tuning.
You can also download the original datasets and process them yourself.
The preprocessing scripts are under the folder preprocess/LC-QuAD2.0-pre.
The configs/train_1.json is an example of parameter configuration for pre-training.
Replace "model_name_or_path" with the model name (t5-small, t5-base, or t5-large) or the path to your checkpoint , "output_dir" with where you want to store your outputs, and "cache_dir" with the place for caching.
You can simply run the code below:
CUDA_VISIBLE_DEVICES=0 python seq2seq/run_seq2seq.py configs/train_1.json
The configs/train_2.json is an example of parameter configuration for fine-tuning the model.
You should replace "dataset" with the name of the dataset that your want to fine-tune the model on, and you can choose from [lc_quad_2, qald_9, qald_10]. Replace "model_name_or_path" with the path to your checkpoint obtained during the previous pre-training.
You can simply run the code below:
CUDA_VISIBLE_DEVICES=0 python seq2seq/run_seq2seq.py configs/train_2.json
