by Google Research / CMU · Natural Language Processing
Extended Transformer architecture that captures much longer dependencies than RNNs through segment-level recurrence and relative positional encoding.
No problems reported for this model.