Sentence-BERT (SBERT) is a modified version of the BERT network that uses siamese and triplet network structures to generate semantically meaningful sentence embeddings. These embeddings can be compared using cosine similarity, significantly reducing the computational effort required to find the most similar pair of sentences. SBERT reduces the time needed for this task from 65 hours with BERT to about 5 seconds, while maintaining the accuracy of BERT. SBERT is evaluated on common semantic textual similarity (STS) tasks and transfer learning tasks, where it outperforms other state-of-the-art sentence embedding methods. It achieves improvements of 11.7 points on seven STS tasks compared to InferSent and 5.5 points compared to Universal Sentence Encoder. On SentEval, SBERT improves by 2.1 and 2.6 points, respectively. SBERT can be adapted to specific tasks and has set new state-of-the-art performance on challenging datasets such as the Argument Facet Similarity (AFS) corpus and a triplet dataset for distinguishing sentences from different sections of a Wikipedia article. SBERT is computationally efficient, with performance on a GPU being about 9% faster than InferSent and 55% faster than Universal Sentence Encoder. It is suitable for tasks that are computationally not feasible with BERT, such as clustering 10,000 sentences, which can be done in about 5 seconds with SBERT instead of 65 hours with BERT. SBERT is trained using a siamese network structure and triplet loss to ensure that sentence embeddings are semantically meaningful. It uses different pooling strategies (MEAN, MAX, and CLS) to derive fixed-sized sentence embeddings. The model is fine-tuned on NLI data, which results in sentence embeddings that significantly outperform other state-of-the-art methods. SBERT is evaluated on various tasks, including STS, AFS, and SentEval, demonstrating its effectiveness in capturing semantic similarity and performing well in transfer learning tasks. The model's efficiency and effectiveness make it a valuable tool for semantic similarity search, clustering, and information retrieval.Sentence-BERT (SBERT) is a modified version of the BERT network that uses siamese and triplet network structures to generate semantically meaningful sentence embeddings. These embeddings can be compared using cosine similarity, significantly reducing the computational effort required to find the most similar pair of sentences. SBERT reduces the time needed for this task from 65 hours with BERT to about 5 seconds, while maintaining the accuracy of BERT. SBERT is evaluated on common semantic textual similarity (STS) tasks and transfer learning tasks, where it outperforms other state-of-the-art sentence embedding methods. It achieves improvements of 11.7 points on seven STS tasks compared to InferSent and 5.5 points compared to Universal Sentence Encoder. On SentEval, SBERT improves by 2.1 and 2.6 points, respectively. SBERT can be adapted to specific tasks and has set new state-of-the-art performance on challenging datasets such as the Argument Facet Similarity (AFS) corpus and a triplet dataset for distinguishing sentences from different sections of a Wikipedia article. SBERT is computationally efficient, with performance on a GPU being about 9% faster than InferSent and 55% faster than Universal Sentence Encoder. It is suitable for tasks that are computationally not feasible with BERT, such as clustering 10,000 sentences, which can be done in about 5 seconds with SBERT instead of 65 hours with BERT. SBERT is trained using a siamese network structure and triplet loss to ensure that sentence embeddings are semantically meaningful. It uses different pooling strategies (MEAN, MAX, and CLS) to derive fixed-sized sentence embeddings. The model is fine-tuned on NLI data, which results in sentence embeddings that significantly outperform other state-of-the-art methods. SBERT is evaluated on various tasks, including STS, AFS, and SentEval, demonstrating its effectiveness in capturing semantic similarity and performing well in transfer learning tasks. The model's efficiency and effectiveness make it a valuable tool for semantic similarity search, clustering, and information retrieval.