A 4-layer Transformer was trained on English constituency parsing (Wall Street Journal portion of the Penn Treebank, ~40K training sentences).
With almost no task-specific tuning, it outperforms the BerkeleyParser — even when trained only on the WSJ training set.
In the semi-supervised setting (with a ~17M-sentence high-confidence corpus) it beats all previously reported models except the Recurrent Neural Network Grammar.
The same architecture generalizes across tasks without structural changes.
91.3 F1
WSJ-only, discriminative setting (Section 23 of WSJ)
92.7 F1
Semi-supervised setting — above every prior model except RNNG
Source: Table 4 (paper p. 10) and §6.2 / “English Constituency Parsing” (paper p. 9–10).