In the context of word embeddings, it's not typical to train words or sentences in two layers with different importance. Word embeddings are usually trained in a single layer where each dimension captures different aspects of the word's meaning, including its syntax, semantics, and context.
However, what you're describing sounds more like a task for a more complex model, such as a transformer model like BERT or a recurrent neural network (RNN) like LSTM. These models can capture both the syntax and semantics of words in a sentence, as well as their position.
If you want to emphasize certain aspects more than others, you could potentially do this during the training process by using different loss functions or by weighting the errors differently. For example, you could assign a higher weight to errors made on the context of the word compared to its syntax or position.