Question regarding kNN-LM (exact)

[jiaqing23]

In the Efficient kNN-LM paper, the result of kNN-LM (exact) is included (with ppl = 16.12, and very slow inference speed). I checked the original kNN-LM (Khandelwal, 2020) and the ppl score is matched with it.

May I know how to you obtain the score for kNN-LM (ppl = 16.65), which part of the code did you change from their original repo (https://github.com/urvashik/knnlm) and make the inference speed increase? What is the different between kNN-LM and kNN-LM (exact)? Because the original kNN-LM paper doesn't have any result of ppl = 16.65.

Thank you very much!

[jxhe]

As their README says, you can set --no-load-keys and --knn-sim-func "do_not_recomp_l2"

[jiaqing23]

Thanks for the response!