Skip to content
Alex Rudnick edited this page Nov 22, 2013 · 1 revision

TreeTransducers

http://tinyurl.com/alexr-quals-transducers

http://tinyurl.com/alexr-quals-volume-1

http://arxiv.org/abs/1203.6136

KURT: http://github.com/alexrudnick/kurt

The Question

What are the tree transducers described in Graehl and Knight's paper "Training Tree Transducers", how are they useful for MT, and are they sufficient to do the kinds of transformations described in Bonnie Dorr's 1994 paper?

major goal: implementing Treen Transducers

  • implement the stuff in Graehl, Knight and May, or at least some of it
  • TODO: get straight answer out of Mike the extent to which it's OK to do three implementation-flavored quals problems

1. given a tree and a transducer: transduce

  • just a search problem, right? keep applying production rules until you're done.
  • maybe do beam search or optionally produce the best ones first or something.

2. given corpus of tree pairs and a set of productions: learn weights

  • EM algorithm for this is given in the paper.

3. How do tree transducers handle... (TODO fill this in) ?

lessons and questions

  • a state is associated with a node in the tree. (ie, each node may have a different state, simultaneously)
  • regular tree grammars
  • what's a Tree Substitution Grammar?
  • how do these relate to CFGs or TAGs?
  • regular tree grammars are a subset of TAGs. TAGs are a superset of CFGs.
  • deep question: regular tree grammars are a superset of CFGs, y/n?

done

  • got pretty good answer from Mike about doing original research for this and the general flavor of the project

CategoryQuals

Clone this wiki locally