Description

Transducers are formal devices for converting between one language and
another. Finite-state transducers, for instance, convert strings in one
regular language to strings in another. They have been used (especially in
their weighted variant) with great success in engineering-oriented
natural-language processing (NLP) applications, such as speech recognition
and morphological processing. Higher-level transducers, such as the many
varieties of tree transducers, have been well studied and applied in the
context of programming language compilation, but have not been applied to
NLP tasks. We will present these formal devices, discussing the design of
transducer devices specifically geared toward NLP tasks. We will place
special emphasis on tree transducers, synchronous grammars, and the
relations between them. The course will culminate in a description of a
"universal natural-language pipeline" based on transducers.