Compare, Compress and Propagate: Enhancing Neural Architectures with Alignment Factorization for Natural Language Inference