It's a machine learning method for turning a sample of text into a fixed-length vector. We usually try to find embeddings with the property that samples of text with similar meanings are mapped to vectors that are close to one another.