Can we perform pseudo-relevance feedback with BM25S?

[roynirmal]

Thanks for the amazing tool!

Is there any way we a perform pseudo-relevance feedback using your implementation? Or any suggestions on how I can do the same?
https://nlp.stanford.edu/IR-book/html/htmledition/pseudo-relevance-feedback-1.html

[xhluca]

Sorry, i'm not familiar with the concept, and it's challenging for me to work out the math wrt non-occurrence matrix and score shifting. If you work it out, feel free to share it here!

Regarding building bm25 scorers in general, you can implement your own idf scorer in a fork and add it here:

bm25s/bm25s/scoring.py

Line 206 in f77e90d

def _select_idf_scorer(method) -> callable:

and create your own tf component scorer and add it here:

bm25s/bm25s/scoring.py

Line 146 in f77e90d

def _select_tfc_scorer(method) -> callable:

then, in the bm25s, you can select it by using the name of your custom bm25 implementation.