Hello, I am thinking in developing an algorithm similar to n-grams, the algorithm should use some recursive procedure to replace terms with context-terms so that the final result of this algorithm would be similar to an embedding, that is a function from n "context words" to probabilities for the next word. Do anyone of you know if this is a solved problem or any attempt in that directions. Thanks in advance.

Context-terms could initially be words but could evolve into something similar to small embeddings.