In our experiments we use the same definition of structural locality as was proposed for the ISBN dependency parser in (Titov and Henderson, 2007b).
We use the neural network approximation (Titov and Henderson, 2007a) to perform inference in our model.
ISBNs, originally proposed for constituent parsing in (Titov and Henderson, 2007a), use vectors of binary latent variables to encode information about the parse history.
Our probabilistic model is based on Incremental Sigmoid Belief Networks (ISBNs), a recently proposed latent variable model for syntactic structure prediction, which has shown very good behaviour for both constituency (Titov and Henderson, 2007a) and dependency parsing (Titov and Henderson, 2007b).
Thus, transition-based parsers normally run in linear or quadratic time, using greedy deterministic search or fixed-width beam search (Nivre et al., 2004; Attardi, 2006; Johansson and Nugues, 2007; Titov and Henderson, 2007), and graph-based models support exact inference in at most cubic time, which is efficient enough to make global discriminative training practically feasible (McDonald et al., 2005a; McDonald et al., 2005b).