04 - Recurrent Neural Network Grammars, with Chris Dyer

Transition Frameworks

Transition frameworks in tree structure generation offer a unique approach by adapting parser transitions to generator transitions. Chris Dyer explains that this involves defining a distribution over structures, allowing trees to be enumerated in various ways. He highlights the importance of factorizing distributions using the chain rule and simplifying structure-building operations into manageable actions.

We're going to generate trees by following this ordering. And this has been done previously in discriminative models.

--- Chris Dyer

This method, rooted in transition-based parsing and abstract state machines, simplifies the complex dependencies across time in tree building 1.

   

Generative Modeling

The generative modeling approach in tree structure generation emphasizes the flexibility of transition-based parsing. Chris Dyer notes that these models can operate on data structures until reaching an end condition, sometimes consuming inputs or generating outputs. He mentions that this approach has been applied to both bottom-up and top-down generation orders, offering diverse possibilities in dependency parsing.

Sometimes some of these automata will consume inputs and other ones will just generate things, or maybe they'll do a mix of the two.

--- Chris Dyer

This flexibility allows for innovative comparisons between different generation orders, particularly in dependency cases 2.