Graph-Induced Syntactic-Semantic Spaces in Transformer-Based Variational AutoEncoders