Partitioning multitape transducers

Abstract

In this paper, we define a class of transducers closed under intersection and complementation, which are the operations used forss contextual rule compilation. This class of transducers is not theoretically more powerful than the Epsilon-Free Letter Transducers most commonly used. But they are more convenient for morphological description whenever the correspondence between lexical and surface forms is not a symbol-to-symbol matching. A complete set of operations on transducers is defined, including some operations (projection and join) which change the number of tapes of the transducers.