A Latent Variable Model for Generative Dependency Parsing
This chapter presents a new version of one of the first latent variable models introduced to the parsing community (Henderson, 2003). The generative dependency parsing model uses binary latent features to induce conditioning features. The induced conditioning features are assumed to be local in the dependency structure, but because induced features are conditioned on other induced features, information can propagate arbitrarily far. The model is formally defined as a recently proposed class of Bayesian Networks for structured prediction, Incremental Sigmoid Belief Networks, and approximated by two methods. The error analysis in this chapter shows that the features induced by the ISBN's latent variables are crucial to this success, and shows that the induced features result in the proposed model being particularly good on long dependencies.