sentence-bert

Definition

A method, also called SBERT, that turns BERT into a sentence embedding model. Vanilla BERT compares two sentences by encoding them jointly, so scoring n sentences against each other costs a quadratic number of forward passes; SBERT trains with a siamese, shared-weight setup so each sentence is encoded once and comparison collapses to a cosine similarity over cached vectors. That is what made semantic search over large corpora practical.