Some Limits on Statistical Learning
Lists
Only you see this
Abstract
Some Limits on Statistical Learning Rebecca Morley ([email protected]) William Badecker ([email protected]) Paul Smolensky ([email protected]) Department of Cognitive Science, Johns Hopkins University 3400 N. Charles St. Baltimore, MD 21218 Introduction VvzV VksV A fairly large body of work within the artificial grammar paradigm has provided some reason to believe that statistical learning may be easy and rapid in both adult and infant subjects (e.g. Newport and Aslin, 2004; Wilson, 2003; Peperkamp et al., 2003). However, there are indications that domain and task may be very critical determiners of learnability. Factors such as degree of explicit knowledge in learning, strength of native language bias, and computational complexity of the data sample seem likely to play a role as well. In the work presented here, subjects were exposed to a lexicon of novel words that exhibited a co-occurrence restriction of some type. Experiment 1 imposed a phonotactic constraint of within cluster voicing assimilation (VCCV word forms). Experiment 2 trained subjects on a (non-explicit) suffix alternation conditioned by the identity of the final stem consonant (CVCV-CV word forms). 2.2 was designed to test the hypothesis that the identity relation (CVzV-zi) may be necessary to bootstrap learning in this domain. CVzV-zi CVvV-zi CVgV-zi CVvV-zi Results & Conclusions Subjects failed to learn in experiments 1.1,1.2, and 2.2, and demonstrated successful learning in experiment 2.1 (for both old and new forms). These results suggest that learning of novel linguistic distributions based on co- occurrence statistics may be more difficult to induce than previously supposed: that the computational device may be quite limited in the number of units it can keep track of, that a highly salient identity relation may be necessary to focus attention on potential structure, etc. Further work will be needed to fully disentangle the contributions of each of these factors. Methods All trials consisted of a pre-test, training, and post-test portion. The pre-test and post-test task was a two- alternative forced choice paradigm which asked subjects to choose which of two auditorily presented tokens was a word in the made-up exposure language (one alternative was grammatical and one ungrammatical according to the particular alternation). Exposure consisted of approximately 20 minutes of auditorily presented word tokens from one of the following data distributions. Acknowledgments This work was supported by an NSF IGERT training grant to the Department, and a Department of Education Javits Fellowship. References Newport,E.L. & Aslin, R.N. (2004). Learning at a Distance I. Statistical Learning of Non-adjacent Dependencies. Cognitive Psychology, 48, 127-162. Peperkamp, S.,Pettinato, M., et al. (2003). Allophonic variation and the acquisition of phoneme categories. BUCLD 27, 650-661. Wilson, C. (2003). Experimental Investigation of Phonological Naturalness. West Coast Conf. Formal Ling. 22, 101-114. Data Distributions VvbV VvgV VvzV VkfV VkpV VksV VbvV VbgV VbzV VfkV VfpV VfsV VgbV VgvV VgzV VpkV VpfV VpsV VbzV VgzV VfsV VpsV CVsV-si CVfV-si CVpV-si CVbV-si CVkV-si CVgV-si CVsV-si CVfV-si CVpV-si CVbV-si CVkV-si VzvV VzbV VzgV VskV VsfV VspV
BibTeX
@inproceedings{badecker2006some,
author = {William Badecker and Rebecca Morley and Paul Smolensky},
title = {{Some Limits on Statistical Learning}},
booktitle = {eScholarship},
year = {2006},
}Loading related content...