Dilution of Embeddings

Success of neural networks has shown that embeddings work. Even though a combination of words can dilute the embeddings, they still work. Even LSA works for document similarity. So it is safe to say that dilution effect is not that much.

But are there better methods??