Sentence-BERT (SBERT) is a BERT-based encoder model designed to represent entire sentences for semantic textual similarity. It processes sentences independently and compares their resulting embeddings with cosine similarity. section 2.3.1
BERT supplies bidirectional contextual representations through masked language modeling. SBERT extends this encoder foundation with siamese and triplet network architectures that emphasize useful sentence embeddings rather than only word-level context. section 2.3.1 section 2.3.1
Sentence similarity
Given two sentences, encode each independently with SBERT, then compare the resulting vectors using cosine similarity to estimate their semantic relationship.


