Skip to content
This repository was archived by the owner on Oct 31, 2023. It is now read-only.

Steps to reproduce LM+FM result in Table 3 #19

Open
tongyao-zhu opened this issue Mar 7, 2023 · 1 comment
Open

Steps to reproduce LM+FM result in Table 3 #19

tongyao-zhu opened this issue Mar 7, 2023 · 1 comment

Comments

@tongyao-zhu
Copy link

Thanks for your great work!

Using the checkpoints you released, I am able to get the result of SEAL (LM+FM, intersective) in Table 3. However, I am wondering how to obtain the result of SEAL (LM, |n| = 5) and SEAL (LM+FM) in the table. There are several flags in SEALSearcher in retrieval.py. Which ones should be set?

In addition, I notice that in retrieval.py, the found_keys are always rescored by rk.rescore_keys. May I know what does this function do? I don't seem to find it in the paper.

Thanks in advance!

@mbevila
Copy link
Contributor

mbevila commented May 5, 2023

rescore_keys simplifies scoring at the expense of compute. Instead of getting the scores directly out of the beam search, we refeed the generated ngrams to the decoder and recompute logprobs. It shouldn't change the performances.

Sign up for free to subscribe to this conversation on GitHub. Already have an account? Sign in.
Labels
None yet
Projects
None yet
Development

No branches or pull requests

2 participants