This is the neural network parser described in
Vaswani & Sagae, 2016. Efficient Structured Inference for Transition-Based Parsing with Neural Networks and Error States. TACL.
The parser uses locally normalized models with error states (see paper above), a straightforward way to perform structured prediction with principled search without the need for full decoding during training.