nyu-dl / dl4marco-bert

BSD 3-Clause "New" or "Revised" License
476 stars 87 forks source link

The task is to do a full rank by BERT, given 1 query. Am I right? #25

Closed guotong1988 closed 5 years ago

guotong1988 commented 5 years ago

Full ranking means scoring all the query-passage pairs. Thank you! @rodrigonogueira4

rodrigonogueira4 commented 5 years ago

Full ranking means scoring all passages in the corpus (~8.8M) as opposed to the top-1000 provided by the MS Marco organizers. The way we do it is the following: Index the 8.8M passages and retrieve the top-100 using BM25. Then we score these top-1000 with BERT.