One-shot and few-shot learning of word embeddings