Reproducing results on QA dataset

[cindy2000sh]

First, thanks authors for opensource the codebase!

I wonder how to reproduce MoL experiments for NQ320K, as I only see configs for sequential retrieval models in configs/ folder, especially for setting-up the training + dataloading code. It would be also helpful if authors can point to some related repositories as well.

[jiaqizhai]

Hi Cindy,

Just saw the message (please email me if something is needed next time).

You can use https://github.com/bailuding/rails/blob/main/rails/similarities/mol/query_embeddings_fns.py#L35 and https://github.com/bailuding/rails/blob/main/rails/similarities/mol/item_embeddings_fns.py#L33 for NQ-specific MoL customizations in the paper. The KNN/ANN related functionalities are common across RecSys/NQ in https://github.com/bailuding/rails/tree/main/rails/indexing.

Our NQ specific training code are built on top of DPR but needs a bit of refactoring before release. I recommend using https://github.com/Hannibal046/nanoDPR (or https://github.com/facebookresearch/DPR) as a starting point for short-term experiment needs.