voidism / EAR

Code for the ACL 2023 long paper - Expand, Rerank, and Retrieve: Query Reranking for Open-Domain Question Answering

Home Page:https://arxiv.org/abs/2305.17080

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

about the GAR best query

yzc111 opened this issue · comments

Hello, I am interested in your work, but I have some questions about your paper, for example, GAR best query which uses the ground truth label. I actually don't know the detail about this. use BM25 will get a similar score between the query expansions with Wikipedia, and then how can I get the best query with ground truth?

Hi! Thanks for the question!

For GAR best query, we did BM25 to retrieve the top-100 passages by each query separately. And then we use the ground truth label (the answer span) to check which passages contain the answer. We find the query that has its top-k passage containing the answer with the smallest k. And it is the GAR best query.
The GAR best query will be considered as the performance upper bound of our method.

Hope the above explanation answers your question!

Yung-Sung