Query about the reranker used in the Multilingual Document Grounded Baseline

[Soham Tiwari]

Hi all,

Can someone help me understand what the re-ranker is doing exactly? What method is used to compute the scores between the passage and the queries? Is the re-ranker even necessary since the ground truth french data only has one positive pid for each sample?

https://github.com/modelscope/modelscope/blob/8abfffc7e52ce97e113042ed0d619d60547f2de3/modelscope/trainers/nlp/document_grounded_dialog_rerank_trainer.py#L181

Yours sincerely,

Soham Tiwari