nreimers
commented
2 years ago Hi @brianhance This is an interesting question for which I don't have a perfect solution yet.
You could check for the relevant results which other close docs exist and return these also to the user. This is similar to a k-NN classifier.
Hi, I would like to implement a feature where given a set of search result, a client can select the most accurate results, and refine the search based off of the selected hits. One idea I had was to do a weighted average of query and selected results, and re-search with the resulting vector. I have also thought about training a model with pairs (query, selected res 1) etc. Or even have the user assign similarity scores to pairs for CE training. Basically I want to allow users to take action that can further improve their results with an SBERT search. Are there any techniques that can achieve this? Thanks.