Improving Textual Network Learning with Variational Homophilic Embeddings