Big question, BERT ranking need triple to train, but...

[guotong1988]

https://msmarco.blob.core.windows.net/msmarcoranking/collectionandqueries.tar.gz
The data above do not contains the negative doc.

@rodrigonogueira4 Thank you!!!

The data is from readme here: https://github.com/nyu-dl/dl4ir-doc2query#ms-marco

[rodrigonogueira4]

To train the se2seq model you only need pairs of queries and relevant documents, hence you don't need negatives.

Please post doc2query-related questions in that repository:
https://github.com/nyu-dl/dl4ir-doc2query

[guotong1988]

But BERT ranking step need the negative doc.

[rodrigonogueira4]

Please note that this repository only contains the code to train doc2query (seq2seq) model. If you want to train BERT re-ranker, please follow the steps in https://github.com/nyu-dl/dl4marco-bert.

Also, please note that training BERT re-ranker on the expanded documents did not give better results than training on the non-expanded (original) documents. I.e., you can use the trained model from https://github.com/nyu-dl/dl4marco-bert to re-rank the expanded documents.

[guotong1988]

Thank you.