Learning reduplication with 2-way finite-state transducers
; Proceedings of The 14th International Conference on Grammatical Inference 2018, PMLR 93:67-80, 2019.
Reduplication is a cross-linguistically common and productive word-formation mechanism. However, there are little to no learning results concerning it. This is partly due to the high computational complexity associated with copying, which often goes beyond standard finite-state technology and partly due to the absence of concrete computational models of reduplicative processes. We show here that reduplication can be modeled accurately and succinctly with 2-way finite-state transducers. Based on this finite-state representation, we identify a subclass of 2-way FSTs based on copying and Output Strictly Local functions. These so-called Concatenated Output Strictly Local functions (C-OSL) can model the majority of attested reduplicative processes we have surveyed. We introduce a simple extension to the inference algorithm for OSL functions that trivially leads to a provably correct learning result for C-OSL functions under the assumption that function concatenation is overtly marked.