Your semantic query expansion raised recall@100 by six...

Semantic expansion often suffers from polarity inversion because context-prediction objectives (like those used in standard word embeddings) score terms based on substitutability rather than relevance. As a result, the model may suggest antonyms or unrelated terms that appear in similar contexts. Simply using a larger model will not fix this, as a better model will only fit the flawed objective more accurately. The diagnostic step is to isolate the expansion terms that caused the click-through drop and analyze their impact on the retrieval score.