An architecture that uses input patterns (like n-grams) to look up learned embeddings, expanding model capacity with minimal computation.