Generative Semantics Text Correspondence Model

We argue that groups of unannotated texts with overlapping and non-contradictory semantics represent a valuable source of information for learning semantic repre- sentations. A simple and efficient infer- ence method …...