What are Sentence Embeddings and how do they differ from word embeddings?

Sentence Embeddings are vector representations of entire sentences that capture their meaning and context. Unlike word embeddings, which encode individual words, sentence embeddings aggregate information from all words in a sentence, considering order and syntax.

Main differences:

• Representation level: word embeddings — words, sentence embeddings — sentences or longer texts.
• Context: sentence embeddings consider the context of the entire sentence, not just individual words.
• Application: sentence embeddings are used for tasks like sentence similarity comparison, semantic search, and text clustering.

Example: word2vec provides a vector for the word "bank," while a sentence embedding for the sentence "I went to the bank to withdraw money" reflects the overall meaning of that sentence.

To obtain sentence embeddings, models like BERT, Universal Sentence Encoder, and others trained at the sentence level are often used.