Understanding the role that prior linguistic experience plays in the acquisition of novel categories is critical to our development of theories of how individuals learn a first or second language.
This dissertation examines learning bias in novel phonetic category learning, as in artificial language learning experiments or second language acquisition. During category learning, learners discover categories largely based on the distribution of their input along phonetic cue dimensions. Learners have been shown to make approximately optimal use of distributional information when discovering categories with single cues, whether or not their native language uses a given cue. In contrast, when learning categories requiring multiple cues, learners frequently attend to only some of the available cues, resulting in poorer category learning. For example, Spanish speakers learning English often show difficulty with the vowel distinction in 'sheep' vs 'ship'. Many learners initially attend primarily to statistically unreliable vowel duration differences although stronger cues are available in the form of spectral differences.
Under the direction of Dr. Lisa Davidson, Sean Christopher Martin will carry out a series of experiments to test the degree to which native language experience gives rise to this type of learning bias in category learning. Native speakers of English will be trained to recognize new vowel categories which are distinguished by pairs of familiar and unfamiliar cues. The stimuli are then systematically manipulated to control which of the cues is more reliable. In this way, the respective contributions of preference for statistically robust cues and bias to attend to cues which the learner's prior experience suggests are reliable can be distinguished.
A computational model of learner behavior accounts for expected results in terms of a hierarchical account of category learning. In order to learn from high-dimensional input like the speech signal, the model develops generalizaitons about cue reliability while learning category structure, giving rise to more efficient first-language learning but later generating learning bias because it initially predicts unfamiliar cues to be low-reliability.
The results of this research will aid understanding of second language learning, highlighting issues which might contribute to poor second language learning outcomes. In addition, the proposed computational model demonstrates an approach which allows for more flexible and adaptable category representations which might allow future speech recognition systems to more easily cope with the natural variability of the speech signal.
