Sentence Embeddings

Sentence embeddings represent whole sentences as vectors so systems can compare meaning, retrieve relevant text, group examples, or support classification. Common approaches use contrastive learning on paraphrases, natural language inference pairs, or translated parallel text, alongside distillation and different pooling methods to build or adapt encoders. Models are evaluated on semantic textual similarity, retrieval, inference, and classification benchmarks, with multilingual and domain-specific tests revealing that strong average scores do not guarantee performance in every language or task. A central tension is that single-vector similarity is efficient, but subtle meaning comparisons can require interactions between sentence pairs; robustness, interpretability, and reliable evaluation remain open challenges.