Reduplication with 2-way Deterministic Finite-STATE Transducers

The theory of finite-state machines underlies many applications in the processing of natural language and speech, including almost all examples of word-formation processes in natural language.

The well-known problematic case is reduplication, a word-formation process found in virtually all natural languages. Broadly speaking, reduplication is process for making words by copying a substring s  from a word w and concatenating them to form s~w  or w~s where ~ marks the boundary between the base w and the reduplicant s. For example, in Indonesian the singular of `book' is pronounced as buku but its plural is pronounced as buku~buku.  The copied substring can be either of a bounded size or unbounded size. This variation in the size of the copy is what makes reduplication challenging.

In this seminar, we discuss the role of reduplication both within theoretical linguistics and computational linguistics, and the challenges it brings. We demonstrate a novel approach which meets these challenges to successfully model virtually all reduplicative processes found in natural language. It relies on an understudied type of
finite-state machinery, two-way deterministic finite-state transducers. 

Since reduplication is ubiquituous in natural language, the incorporation of two-way finite-state transducers into existing applications has the potential to bring speech and language processing applications to a whole new level.