Official code for the paper titled Augmenting Transformers with Recursively Composed Multi-grained Representations

In this work, we successfully combine a composition model with bi-directional Transformers and make them jointly pre-trainable.