Lam, C. D.
The University of Chicago
Statistical learning has pervaded different fields of language sciences but its applications in historical linguistics remain limited. Here, I investigate the influence of statistical patterns on the emergence of determiner-noun fusion (DNF) in Haitian Creole (HC). DNF is when a determiner-noun combination in French becomes reinterpreted as a noun with the same meaning in the Creole (HC lapli from FR la pluie, the rain). Most accounts have looked at L1 bias from Bantu languages spoken by slave populations (Baker 1984). However, I hypothesize that statistical patterns in French, including gross and co-occurrence frequencies, and backwards transitional probability (Pelucchi et al., 2009) between the determiner and the noun can also predict the DNF patterns in HC. All three parameters are higher (unpaired t-tests, p<0.0005) in fused nouns (N=259) than in stratified-sampled unfused nouns (N=129). In generalized linear mixed model, co-occurrence frequencies and backwards transitional probability are significant. 100-fold cross validation with these parameters yields a mean accuracy of 79%. Thus, statistical learning is important in DNF emergence and it is possible that DNF results from word boundary misanalysis due to French statistical patterns. This shows that statistical learning can be used to explain emergence of novel patterns in world languages.