Our simplest approach turned out to be our most effective. By removing the less important query
terms, we are amplifying the effects of the SPLADE
log-saturated weighting. That is, we are making the most important terms even more important in
determining the relevance of candidate passages.
On top of that, this best leverages the existing inverted index and search approach. So there’s
very little work required to implement it.
The best version of this is probably reducing the SPLADE Terms down to ⅕ its original size.
Although reducing down to ½ has virtually no effect on the
quality of the results, while significantly speeding up retrieval time.