Towards Probabilistic Acceptors and Transducers for Feature Structures.

Weighted finite-state acceptors and transducers (Pereira and Riley, 1997) are a critical technology for NLP and speech systems. They flexibly capture many kinds of stateful left-to-right substitution, simple transducers can be composed into more complex ones, and they are EM- trainable. They are unable to handle long-range syntactic movement, but tree acceptors and transducers address this weakness (Knight and Graehl, 2005). Tree automata have been profitably used in syntax-based MT systems. Still, strings and trees are both weak at representing linguistic structure involving semantics and reference ("who did what to whom"). Feature structures provide an attractive, well-studied, standard format (Shieber, 1986; Rounds and Kasper, 1986), which we can view computationally as directed acyclic graphs. In this paper, we develop probabilistic acceptors and transducers for feature structures, demonstrate them on linguistic problems, and lay down a foundation for semantics-based MT.