Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks

TL;DR

Sentence-BERT (SBERT) modifies BERT with siamese and triplet networks to generate efficient, high-quality sentence embeddings suitable for semantic similarity tasks, drastically reducing computation time while maintaining accuracy.

Contribution

The paper introduces SBERT, a novel BERT-based architecture that produces semantically meaningful sentence embeddings suitable for fast similarity search.

Findings

SBERT reduces similarity computation time from 65 hours to 5 seconds.

SBERT outperforms existing sentence embedding methods on STS and transfer learning tasks.

SBERT maintains BERT-level accuracy while enabling efficient semantic similarity search.

Abstract

BERT (Devlin et al., 2018) and RoBERTa (Liu et al., 2019) has set a new state-of-the-art performance on sentence-pair regression tasks like semantic textual similarity (STS). However, it requires that both sentences are fed into the network, which causes a massive computational overhead: Finding the most similar pair in a collection of 10,000 sentences requires about 50 million inference computations (~65 hours) with BERT. The construction of BERT makes it unsuitable for semantic similarity search as well as for unsupervised tasks like clustering. In this publication, we present Sentence-BERT (SBERT), a modification of the pretrained BERT network that use siamese and triplet network structures to derive semantically meaningful sentence embeddings that can be compared using cosine-similarity. This reduces the effort for finding the most similar pair from 65 hours with BERT / RoBERTa to about 5 seconds with SBERT, while maintaining the accuracy from BERT. We evaluate SBERT and SRoBERTa on common STS tasks and transfer learning tasks, where it outperforms other state-of-the-art sentence embeddings methods.