Combining Embeddings to Produce New Representative Embedding

I figured I’d put this out there in case miracles can happen. Given two or more embeddings, is it possible to construct a new embedding that is the semantic averaging of the input embeddings? For example, I want to take the embeddings for “I love baseball” and “I like baseball” and come up with an embedding that basically means “I like baseball a lot”.


I was thinking on similar lines as well. Given an embedding, is it possible to retrieve the word associated with it?