Representational Bias in Unsupervised Learning of Syllable Structure

Unsupervised learning algorithms based on Expectation Maximization (EM) are often straightforward to implement and provably converge on a local likelihood maximum. However, these algorithms often do not perform well in practice. Common wisdom holds that they yield poor results because they are overly sensitive to initial parameter values and easily get stuck in local (but not global) maxima. We present a series of experiments indicating that for the task of learning syllable structure, the initial parameter weights are not crucial. Rather, it is the choice of model class itself that makes the difference between successful and unsuccessful learning. We use a language-universal rule-based algorithm to find a good set of parameters, and then train the parameter weights using EM. We achieve word accuracy of 95.9% on German and 97.1% on English, as compared to 97.4% and 98.1% respectively for supervised training.