Efficient finite state unification morphology

Abstract

Finite state transducers are highly efficient means for the representation and processing of morphological knowledge. However, the string representations normally used do not easily provide the rich and detailed linguistic descriptions needed for complex applications in Computational Linguistics, and long-distance phenomena are not easily modeled. This paper describes the use of typed feature structure as weights on transitions in a finite state transducer to represent linguistic objects. This method provides a seamless integration into other linguistic processing modules and facilitates the description of certain morphological phenomena. By using a pre-computation model of unification, we avoid the runtime complexity of unification and achieve a level of efficiency comparable to character-based automata based on other weight structures.