# Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks

**Authors:** Nils Reimers, Iryna Gurevych  
**arXiv:** [1908.10084](https://arxiv.org/abs/1908.10084)  
**Date:** 2019-08-27  
**Summary page:** https://www.summarizepaper.com/en/paper/1908.10084v1/

## Key points

- Sentence-BERT (SBERT) is a modification of the BERT network
- SBERT utilizes siamese and triplet network structures to derive semantically meaningful sentence embeddings
- SBERT allows for efficient semantic similarity search and unsupervised tasks like clustering without requiring both sentences to be fed into the network
- SBERT significantly reduces computational overhead for finding similar pairs in a collection of 10,000 sentences from 65 hours with BERT to just 5 seconds while maintaining accuracy
- SBERT outperforms other state-of-the-art sentence embedding methods in various sentence-pair regression tasks and transfer learning tasks
- SBERT enables BERT to be applied to new tasks that were previously not feasible
- SBERT achieves improved performance in several NLP applications.

## Summary

The paper introduces Sentence-BERT (SBERT), a modification of the BERT network that utilizes siamese and triplet network structures to derive semantically meaningful sentence embeddings. SBERT allows for efficient semantic similarity search and unsupervised tasks like clustering without requiring both sentences to be fed into the network. The authors demonstrate that SBERT significantly reduces the computational overhead required for finding the most similar pair in a collection of 10,000 sentences from 65 hours with BERT to just 5 seconds while maintaining accuracy. They evaluate SBERT on various sentence-pair regression tasks and transfer learning tasks where it outperforms other state-of-the-art sentence embedding methods. Overall, SBERT enables BERT to be applied to new tasks that were previously not feasible and achieves improved performance in several NLP applications.