Why are dense embeddings better than one$-$hot vectors or sparse distributional similarity vectors as input to neural networks?

Choice $1$ of $4$:because embeddings are more modernChoice $2$ of $4$:because embeddings allow us to use networks with fewer parametersChoice $3$ of $4$:because neural networks cannot handle zeros as inputChoice $4$ of $4$:because dense embeddings capture more information than distributional similarities