Modelling the formation of phonotactic restrictions across the mental lexicon

Experimental data shows that adult learners of an artificial language with a phonotactic restriction learned this restriction better when being trained on word types (e.g. when they were presented with 80 different words twice each) than when being trained on word tokens (e.g. when presented with 40 different words four times each) (Hamann & Ernestus submitted). These findings support Pierrehumbert's (2003) observation that phonotactic co-occurrence restrictions are formed across lexical entries, since only lexical levels of representation can be sensitive to type frequencies.
We present a computational model that can explain the type- vs. token learning effects, namely bidirectional phonology and phonetics (Boersma 2007) with an additional semantic level (Apoussidou 2007).