Trained model on MetaQA?

[ianbstewart]

I'm interested in testing the implementation of the text-to-Cypher model, and it seems like that corresponds to the MetaQA dataset. Do you have an available model checkpoint for the text-to-Cypher model?

[Flitternie]

Thanks for your interest in our work! Specifically, in our work, text-to-Cypher conversion is done in 2 steps:

1. text-to-GraphQ IR: as for this part, we have released the MetaQA-Cypher annotation and you may use it for fine-tuning your own checkpoints based on any pretrained language models (we used BART-base in our paper). Since GraphQ IR is a unified IR for multiple graph query languages, you are recommended to make use of other annotation resources (e.g., KQA Pro, GrailQA, or other SPARQL-annotated datasets) to further pretrain the PLM prior to your downstream fine-tuning in order to achieve better end-to-end performance.
2. GraphQ IR-to-Cypher: for this part, we have open-sourced the GraphQ Trans, which is a source-to-source compiler that supports lossless transpilation from GraphQ IR to Cypher. Please see: GraphQ_Trans.

Please feel free to raise any follow-up questions.

[ianbstewart]

Thanks for the information. Did you save the BART-base models that you trained in your study, and if so are those available for download?

[Flitternie]

Unfortunately, we cannot provide public access to the fine-tuned checkpoints. Please consult the paper for the fine-tuning process and let me know if you have any questions.