Modern language–based applications rely on dense vector representations—embeddings—that capture the semantic and syntactic relationships between words far more effectively than one-hot encodings. This chapter begins by defining the embedding layer as a learnable lookup table that maps discrete token IDs into continuous high-dimensional vectors, and shows how these embeddings are trained end-to-end via backpropagation. We then distinguish between global embedding models (Word2Vec, GloVe, FastText) that assign a single vector per word, and contextual models (ELMo, BERT, GPT) that produce different vectors depending on each word’s surrounding text. To illustrate practical usage, we implement both a skip-gram Word2Vec model—with negative sampling—to learn global embeddings, and a many-to-one RNN classifier that jointly learns contextual embeddings while categorizing news topics. Finally, we explore how to visualize and compare learned vectors via cosine similarity and t-SNE and examine how these embeddings can power downstream tasks such as content recommendation.