RETINAQA

• RETINAQA is a robust KBQA model for detecting unanswerability.
• It is a multi-staged Retrieve, Generate and Rank framework.
• It contains major 3 components : Sketch Generation, Logical Form Integrator and a Discriminator.
• Logical Form Integrator combines outputs of sketch generator and two types of reteriver - schema reteriver and path reteriver (similar to TIARA).

Environment Setup

conda create -n retinaqa python=3.9
pip install -r requirements.txt

Training RetinaQA from scratch

1. Run setup.py to download data and retriever outputs for different datasets.

   • If you want to re-train all retriever modules - entity linker, schema reteriver and logical form reteriver, then follow steps mentioned in https://github.com/microsoft/KC/tree/main/papers/TIARA.
   • Copy the output files - entities, top-k relations, top-k classes and top-k paths to appropriate location inside ./data (Example given in setup.py file)

2. Train sketch generator : Follow steps as mentioned in src/sketch_generation.

3. Generate candidate logical forms : Follow steps as mentioned in src/lf_integrator

4. Train discriminator : Follow steps as mentioned in src/discriminator

Model Checkpoints

To directly run inference on GrailQability please find the links to pre-trained model checkpoints for different stages here. Please follow instructions in respective sections to run inference.

Training Type	Module	Checkpoint link
A	Sketch Generation Logical Form Candidates Cache Discriminator	link link link
AU	Sketch Generation Logical Form Candidates Cache Discriminator	link link link

Acknowledgements

Parts of our code/data are taken from:

• https://github.com/dair-iitd/GrailQAbility
• https://github.com/microsoft/KC/tree/main/papers/TIARA
• https://github.com/salesforce/rng-kbqa/