Learning a phonologically partitioned lexicon

Some languages show selective phonological processes whose application is limited to a lexical subgroup. Within the same language, a given lexical item is subject to such a process or exempt from it depending on its subgroup membership. Such subgroups are often defined in etymological terms. That is, words of the same historical origin pattern together with respect to whether they undergo a selective process. However, given that speakers do not have direct access to diachronic information, purely etymological accounts face a conceptual challenge. This dissertation instead tests whether phonotactics can serve as a synchronic phonological factor that partitions the lexicon for selective process application. This study approaches this issue through an investigation of L-Tensification (LT) in Korean using nonce-word experiments. LT is a selective process in which /l/ tensifies a following lenis coronal obstruent. Previous studies have described its applicability as limited to Sino-Korean words. In the experiments, 144 nonce words were used. They were divided into three groups of 48: one with phonotactic cues associated with the Sino-Korean class, one with cues associated with non-Sino-Korean words, and one group of phonotactically neutral words without any such cues. In the production experiment, 42 native speakers of Korean participated in Seoul, and their productions of these nonce words were measured for longer unvoiced duration and closure duration, which are acoustic correlates of tense obstruents in Korean. Separately, a Transformer machine-learning model was trained on a Korean corpus, and its predictions for the same nonce items were analyzed for confidence in whether LT should apply. The results suggest that both speakers and the Transformer model applied LT selectively to nonce words, and that this selectivity was sensitive to phonotactic manipulations. The specific manifestation of the phonotactic effect in LT, however, differed between speakers and the model. Based on these results, the dissertation suggests a possibility that phonotactics can provide a synchronic basis for lexicon partitioning relevant to selective process application. It also contributes to a better understanding of LT and the learnability of selective processes.