It seems like the new standard in text search is sentence or document similarity, using things like BERT sentence embeddings. However, these don't really have a way to consider the salience of sentences, which can make it hard to compare different searches.
For example, when using concept embeddings I'd like to be able to score "Exam" <-> "Exam" as less important than "Diabetes" <-> "High blood sugar". But obviously, the former has a similarity score of 1.
I've tried using weighting with inverse document frequency of terms, but with the latter being unbounded, it's really hard to figure out how to weight similarity and frequency scores. It's also just not as sophisticated, since it doesn't account for synonymy.
Has anybody come up with any solutions to this?