Identity Avoidance Between Non-Adjacent Consonants in Artificial Language Segmentation
Natalie Boll‐Avetisyan, René Kager · 2008
Phonotactic distributions over lexicons correlate with human performance (Coleman & Pierrehumbert, 1997; Vitevitch & Luce, 1999; Pitt & McQueen, 1998). Although it is known that phonotactic probability is independent of lexical analogy (Bailey & Hahn, 2001), the issue remains whether gradient distributions over the lexicon are represented in the grammar in the form of abstract constraints involving natural classes (Frisch, Pierrehumbert, & Broe, 2004; Hayes & Wilson, 2007). This paper supports the hypothesis that gradient distributions over the lexicon may be represented as abstract constraints, which affect segmentation. We tested this hypothesis using artificial language segmentation, a task which is known to be affected by phonological properties of the native language (Onnis, Monaghan, Chater, & Richmond, 2005). We focussed on non-adjacent C_C dependencies, allowing us to test effects of abstract constraints while controlling for phonotactic probabilities of the stimulus material. It has been found that statistical calculations over triphones (e.g. CVC) are uneffective for nonword processing (Bailey & Hahn, 2001). Yet artificial language learning studies show that statistical dependencies between non-adjacent consonants are learnable (Newport & Aslin, 2004). Many languages restrict the co-occurrence of homorganic consonants across intervening vowels (e.g. Frisch & Zawaydeh 2001). Artificial language studies and cross-linguistic studies are consistent with the prediction that constraints on non-adjacent consonants with a probabilistic basis in the lexicon affect artificial language segmentation.