Recursive Non-Autoregressive Graph-to-Graph Transformer for Dependency Parsing with Iterative Refinement

We propose the Recursive Non-autoregressive Graph-to-Graph Transformer architecture (RNGTr) for iterative refinement of arbitrary graphs through recursive application a non-autoregressive and apply it to syntactic dependency parsing. demonstrate power effectiveness RNGTr on several corpora, using model pre-trained with BERT. also introduce Syntactic (SynTr), non-recursive parser similar our model. can improve accuracy variety initial parsers 13 languages from Universal Dependencies Treebanks, English Chinese Penn German CoNLL2009 corpus, even improving over new state-of-the-art results achieved by SynTr, significantly all corpora tested.