Feature-driven new sound category formation: computational implementation with the L2LP model and beyond

Abstract

One of the primary questions of second language (L2) acquisition research is how a new sound category is formed to allow for an L2 contrast that does not exist in the learner's first language (L1). Most models rely crucially on perceived (dis)similarities between L1 and L2 sounds, but a precise definition of what constitutes “similarity” has long proven elusive. The current study proposes that perceived cross-linguistic similarities are based on feature-level representations, not segmental categories. We investigate how L1 Japanese listeners learn to establish a new category for L2 American English /æ/ through a perception experiment and computational, phonological modeling. Our experimental results reveal that intermediate-level Japanese learners of English perceive /æ/ as an unusually fronted deviant of Japanese /a/. We implemented two versions of the Second Language Linguistic Perception (L2LP) model with Stochastic Optimality Theory—one mapping acoustic cues to segmental categories and another to features—and compared their simulated learning results to the experimental results. The segmental model was theoretically inadequate as it was unable explain how L1 Japanese listeners notice the deviance of /æ/ from /a/ in the first place, and was also practically implausible because the predicted overall perception patterns were too native English-like compared to real learners' perception. The featural model, however, showed that the deviance of /æ/ could be perceived due to an ill-formed combination of height and backness features, namely */low, front/. The featural model, therefore, reflected the experimental results more closely, where a new category was formed for /æ/ but not for other L2 vowels /ɛ/, /ʌ/, and /ɑ/, which although acoustically deviate from L1 /e/, /a/, and /o/, are nonetheless featurally well-formed in L1 Japanese, namely /mid, front/, /low, central/, and /mid, back/. The benefits of a feature-based approach for L2LP and other L2 models, as well as future directions for extending the approach, are discussed.