Biases in segmenting non-concatenative morphology

Abstract

Segmentation of words containing non-concatenative morphology into their component morphemes, such as Arabic /kita:b/ 'book' into root [check symbol]ktb and vocalism /i-a:/ (McCarthy, 1979, 1981), is a difficult task due to the size of its search space of possibilities, which grows exponentially as word length increases, versus the linear growth that accompanies concatenative morphology. In this dissertation, I investigate via computational and typological simulations, as well as an artificial grammar experiment, the task of morphological segmentation in root-and-pattern languages, as well as the consequences for majority-concatenative languages such as English when we do not presuppose concatenative segmentation and its smaller hypothesis space. In particular, I examine the necessity and sufficiency conditions of three biases that may be hypothesised to govern the learning of such a segmentation: a bias towards a parsimonious morpheme lexicon with a power-law (Zipfian) distribution over tokens drawn from this lexicon, as has successfully been used in Bayesian models of word segmentation and morphological segmentation of concatenative languages (Goldwater et al., 2009; Poon et al., 2009, et seq.); a bias towards concatenativity; and a bias against interleaving morphemes that are mixtures of consonants and vowels. I demonstrate that while computationally, the parsimony bias is sufficient to segment Arabic verbal stems into roots and residues, typological considerations argue for the existence of biases towards concatenativity and towards separating consonants and vowels in root-and-pattern-style morphology. Further evidence for these as synchronic biases comes from the artificial grammar experiment, which demonstrates that languages respecting these biases have a small but significant learnability advantage.