triplet-loss-from-facenet

IN premise — summaries/2026/08/24/reimers-2019-sentence-bert-sR-references.md

Created 2026-08-25T02:58:30+00:00

Sentence-BERT's triplet loss formulation (minimize distance(anchor, positive) − distance(anchor, negative) + margin) is borrowed from FaceNet (Schroff et al., 2015).

Summary

Sentence-BERT learns to embed sentences by pulling matching pairs closer together and pushing non-matching pairs apart, a training strategy originally designed for face recognition. Knowing this lineage matters because it reveals that the geometric and margin-based design choices were optimized for a different domain, which shapes how well the model handles the subtleties of natural language semantics.