Bi-LSTM part of the neural architecture. Each word is represented by the concatenation of a standard word-embedding w and the output of a character bi-LSTM c. The concatenation is fed to a two-layer bi-LSTM transducer that produces contextual word representations. The first layer serves as input to the tagger (Section 4.2), whereas the second layer is used by the parser to instantiate feature templates for each parsing step (Section 4.3).